Academic Journal
View in EDS
PDF Full Text
HTML Full Text

Software BreastAnalyser for the semi-automatic analysis of breast cancer immunohistochemical images

Bibliographic Details
Title: Software BreastAnalyser for the semi-automatic analysis of breast cancer immunohistochemical images
Authors: Marina Rodríguez-Candela Mateos, Maria Azmat, Paz Santiago-Freijanes, Eva María Galán-Moya, Manuel Fernández-Delgado, Rosa Barbella Aponte, Joaquín Mosquera, Benigno Acea, Eva Cernadas, María D. Mayán
Source: Scientific Reports, Vol 14, Iss 1, Pp 1-19 (2024)
Publisher Information: Nature Portfolio, 2024.
Publication Year: 2024
Collection: LCC:Medicine
LCC:Science
Subject Terms: Medicine, Science
More Details: Abstract Breast cancer is the most diagnosed cancer worldwide and represents the fifth cause of cancer mortality globally. It is a highly heterogeneous disease, that comprises various molecular subtypes, often diagnosed by immunohistochemistry. This technique is widely employed in basic, translational and pathological anatomy research, where it can support the oncological diagnosis, therapeutic decisions and biomarker discovery. Nevertheless, its evaluation is often qualitative, raising the need for accurate quantitation methodologies. We present the software BreastAnalyser, a valuable and reliable tool to automatically measure the area of 3,3’-diaminobenzidine tetrahydrocholoride (DAB)-brown-stained proteins detected by immunohistochemistry. BreastAnalyser also automatically counts cell nuclei and classifies them according to their DAB-brown-staining level. This is performed using sophisticated segmentation algorithms that consider intrinsic image variability and save image normalization time. BreastAnalyser has a clean, friendly and intuitive interface that allows to supervise the quantitations performed by the user, to annotate images and to unify the experts’ criteria. BreastAnalyser was validated in representative human breast cancer immunohistochemistry images detecting various antigens. According to the automatic processing, the DAB-brown area was almost perfectly recognized, being the average difference between true and computer DAB-brown percentage lower than 0.7 points for all sets. The detection of nuclei allowed proper cell density relativization of the brown signal for comparison purposes between the different patients. BreastAnalyser obtained a score of 85.5 using the system usability scale questionnaire, which means that the tool is perceived as excellent by the experts. In the biomedical context, the connexin43 (Cx43) protein was found to be significantly downregulated in human core needle invasive breast cancer samples when compared to normal breast, with a trend to decrease as the subtype malignancy increased. Higher Cx43 protein levels were significantly associated to lower cancer recurrence risk in Oncotype DX-tested luminal B HER2- breast cancer tissues. BreastAnalyser and the annotated images are publically available https://citius.usc.es/transferencia/software/breastanalyser for research purposes.
Document Type: article
File Description: electronic resource
Language: English
ISSN: 2045-2322
Relation: https://doaj.org/toc/2045-2322
DOI: 10.1038/s41598-024-53002-6
Access URL: https://doaj.org/article/231a556edeec4e41961f4e706f43b3cc
Accession Number: edsdoj.231a556edeec4e41961f4e706f43b3cc
Database: Directory of Open Access Journals
Full text is not displayed to guests.
FullText Links:
  – Type: pdflink
    Url: https://content.ebscohost.com/cds/retrieve?content=AQICAHjPtM4BHU3ZchRwgzYmadcigk49r9CVlbU7V5F6lgH7WwGivDiEpux8qQxBx7DFP_9NAAAA4jCB3wYJKoZIhvcNAQcGoIHRMIHOAgEAMIHIBgkqhkiG9w0BBwEwHgYJYIZIAWUDBAEuMBEEDCRg5dchEw-gdc9_wAIBEICBmjyPMImnFuk_Vhv3ydt5OSGy2llF5tvHQw3ASjBhk3uC0afrglfsOQRnJfvyUja09XGqCRvk087L24VPOs-GOO1AFe8ui2tN_GbWKd0uSMY6VdwaIaXRQmNy3E6VcztzC1nvx1IWgwlWWtDpD61eXCTW9PXTSkl0v3LQ6vOQ2rxD-3no35vYKn8oue1jFJL_RLT0Wv1U2bdDlBA=
Text:
  Availability: 1
  Value: <anid>AN0175756719;[fkqs]06feb.24;2024Mar04.06:35;v2.2.500</anid> <title id="AN0175756719-1">Software BreastAnalyser for the semi-automatic analysis of breast cancer immunohistochemical images </title> <p>Breast cancer is the most diagnosed cancer worldwide and represents the fifth cause of cancer mortality globally. It is a highly heterogeneous disease, that comprises various molecular subtypes, often diagnosed by immunohistochemistry. This technique is widely employed in basic, translational and pathological anatomy research, where it can support the oncological diagnosis, therapeutic decisions and biomarker discovery. Nevertheless, its evaluation is often qualitative, raising the need for accurate quantitation methodologies. We present the software BreastAnalyser, a valuable and reliable tool to automatically measure the area of 3,3'-diaminobenzidine tetrahydrocholoride (DAB)-brown-stained proteins detected by immunohistochemistry. BreastAnalyser also automatically counts cell nuclei and classifies them according to their DAB-brown-staining level. This is performed using sophisticated segmentation algorithms that consider intrinsic image variability and save image normalization time. BreastAnalyser has a clean, friendly and intuitive interface that allows to supervise the quantitations performed by the user, to annotate images and to unify the experts' criteria. BreastAnalyser was validated in representative human breast cancer immunohistochemistry images detecting various antigens. According to the automatic processing, the DAB-brown area was almost perfectly recognized, being the average difference between true and computer DAB-brown percentage lower than 0.7 points for all sets. The detection of nuclei allowed proper cell density relativization of the brown signal for comparison purposes between the different patients. BreastAnalyser obtained a score of 85.5 using the system usability scale questionnaire, which means that the tool is perceived as excellent by the experts. In the biomedical context, the connexin43 (Cx43) protein was found to be significantly downregulated in human core needle invasive breast cancer samples when compared to normal breast, with a trend to decrease as the subtype malignancy increased. Higher Cx43 protein levels were significantly associated to lower cancer recurrence risk in Oncotype DX-tested luminal B HER2- breast cancer tissues. BreastAnalyser and the annotated images are publically available https://citius.usc.es/transferencia/software/breastanalyser for research purposes.</p> <p>These authors contributed equally: Marina Rodríguez-Candela Mateos and Maria Azmat.</p> <hd id="AN0175756719-2">Introduction</hd> <p>According to the latest GLOBOCAN estimates for 2020[<reflink idref="bib1" id="ref1">1</reflink>],[<reflink idref="bib2" id="ref2">2</reflink>], breast carcinoma is the most diagnosed cancer in both sexes worldwide (11.7%). It also accounts for the fifth cause of cancer mortality globally (6.9%) and the highest cancer-related mortality among women (15.5%). Breast cancer is a highly heterogeneous disease, comprising various profiles with different histopathological features detected by immunohistochemistry (IHC) analysis according to intrinsic molecular traits, clinical behaviours, and treatment responses. Traditional stratification follows the initial studies of Perou et al.[<reflink idref="bib3" id="ref3">3</reflink>] and Sørlie et al.[<reflink idref="bib4" id="ref4">4</reflink>], leading to the acknowledgment of five intrinsic molecular subtypes: luminal A, luminal B, HER2 overexpressing, basal and normal-like tumours. Luminal A breast tumours express estrogen receptor (ER) and/or progesterone receptor (PR), are negative for human epidermal growth factor receptor 2 (HER2), have low levels of protein Ki67 and are usually of grade 1 or 2. Their luminal B counterparts also express hormone receptor/s but some are positive for HER2; they are usually of higher grade (2–3) and express higher levels of Ki67 and other proliferation-related genes. Both luminal subtypes render the greatest outcome, although luminal A are linked to a significantly better prognosis. Both respond well to endocrine therapy and represent the most common breast cancer subtypes. HER2 enriched tumours are negative for hormone receptors and overexpress HER2 (<emph>ERBB2</emph>) and other genes in its amplicon such as <emph>GRB7</emph> and <emph>PGAP3</emph>. They are generally of high grade (2–3) and result in a poor outcome. They respond to available targeted agents, such as trastuzumab, an anti-HER2 monoclonal antibody. Basal-like/triple negative tumours show low expression or lack hormone receptors and HER2, and some patients/tumours express higher levels of proliferation-related genes and basal markers like keratins 5, 6, 14, 17 and epidermal growth factor receptor (EGFR). Usually of grade 3, they are defined by their aggressiveness, poorest prognosis, higher risk of relapse and metastasis, and lower disease-specific survival. This subtype has currently no targeted therapies available. Normal-like breast tumours share similar pathological markers as luminal A (ER+, PR+, HER2-, Ki67 low), show a normal breast tissue profiling, and result in an intermediate prognosis[<reflink idref="bib3" id="ref5">3</reflink>]–[<reflink idref="bib5" id="ref6">5</reflink>]. Regarding to this matter, it is important to draw attention to the grounbreaking multigene signature-based tests, such as Oncotype DX® (ODX; Exact Sciences, Madison, WI) that help guide clinical treatment decisions in ER+ HER2- lymph node negative early breast cancer[<reflink idref="bib6" id="ref7">6</reflink>]. A highly cost-effective test, it has been widely and clinically validated and it is the only gene-based assay certified to predict prognosis and benefit from chemotherapy treatment at the same time. Patients with scores between 0 and 25 present lower cancer recurrence risk if they undergo hormonal treatment, but are less likely to benefit from chemotherapy. Conversely, scores 26 and higher are indicative of high risk of cancer recurrence with hormonal therapy, whereas these patients can better benefit from adjuvant chemotherapy[<reflink idref="bib7" id="ref8">7</reflink>]–[<reflink idref="bib9" id="ref9">9</reflink>].</p> <p>In the context of breast anatomo-histopathology, 3,3'-diaminobenzidine tetrahydrochloride (DAB)-based IHC is a technique that stains specific protein antigens in brown, usually employing a cellular counterstain, such as hematoxylin, which renders the nuclei purplish blue. It has been regarded as a low cost yet highly complementary methodology, aiding in the diagnosis, subtyping and therapeutic indications of neoplasias. Another relevant application of IHC is related to the search of prognostic factors and biomarkers[<reflink idref="bib10" id="ref10">10</reflink>]. However, IHC assessment is often performed qualitatively and subjectively, such as presence/absence of target antigen, biasing its interpretation. Furthermore, apart from the degree of expression of a specific molecule, it is also crucial to discriminate its localization in various cellular compartments, such as the nuclei, which might be indicative of different biological roles[<reflink idref="bib11" id="ref11">11</reflink>]. Subsequently, there is a growing need for more accurate and reliable methods for IHC quantitation, with computer-based image analysis resulting in higher precision, solidity and quality in IHC quantification[<reflink idref="bib12" id="ref12">12</reflink>].</p> <p>Among the most popular tools in the biomedical field to analyse and quantify IHC images we find ImageJ[<reflink idref="bib13" id="ref13">13</reflink>]. It provides many common image processing algorithms and allows the definition of customized processing plugins. However, ImageJ does not allow object outlines to be corrected manually in a versatile and easy way before starting the image quantification. It is also worth mentioning Qupath[<reflink idref="bib14" id="ref14">14</reflink>], an open source software for digital pathology image analysis. For DAB-IHC quantitation, Qupath offers a wide array of parameters that need to be fine-tuned in order to improve signal detection, resulting in a time-consuming process that requires certain expertise knowledge. Other recent approaches[<reflink idref="bib15" id="ref15">15</reflink>]–[<reflink idref="bib21" id="ref16">21</reflink>] process automatically the IHC or histological images to detect and/or classify interest objects in the image, but they do not allow any expert supervision before the quantification, or require specific and expensive devices. We developed the CystAnalyser[<reflink idref="bib22" id="ref17">22</reflink>] and STERapp[<reflink idref="bib23" id="ref18">23</reflink>] software tools to quantitatively analyse histological images in medicine and biology, respectively. These tools overcome some limitations of ImageJ, QuPath and other approaches through a friendly graphical interface (GUI) easier to use for biomedical experts. Both softwares use image analysis and machine learning algorithms to automatically recognise and classify the objects of interest in the image, allowing the experts to review the recognition of objects using the GUI before measuring and counting them.</p> <p>Graph: Figure 1Representative IHC images against Cx43 (brown) and hematoxylin nuclear counter-stain (purple) of core needle biopsies of invasive breast cancer tumours not subjected to chemotherapy or radiotherapy, belonging to all 6 breast cancer subtypes: healthy mammary tissue, luminal A BC, luminal B HER2+ BC, luminal B HER2- BC, HER2+ BC and triple negative BC.</p> <p>This paper proposes the software BreastAnalyser to quantify breast cancer immunohistochemical images of antigens relevant in breast anatomo-histopathology. In this work we focus, as a representative example, on the staining of the gap junction protein connexin43 (Cx43), in order to compare Cx43 levels among different subtypes and various degrees of breast cancer malignancy and risk of recurrence. Connexins (Cxs) are the basic protein components of the gap junction (GJ) channels and hemichannels, present in most cells and tissues. GJs allow direct communication between the cytoplasms of neighbouring cells, leading to the bidirectional passage of electrical signals and most small and soluble second messengers and molecules[<reflink idref="bib24" id="ref19">24</reflink>]. Connexins can also regulate several signalling pathways by their interaction with different protein partners[<reflink idref="bib25" id="ref20">25</reflink>]. Cx43 has been identified as a crucial component during mammary gland epithelial differentiation and development[<reflink idref="bib26" id="ref21">26</reflink>]. Loss of Cx43 and GJ functionality has been widely reported in breast tumour cell lines[<reflink idref="bib27" id="ref22">27</reflink>],[<reflink idref="bib28" id="ref23">28</reflink>] and primary tumours[<reflink idref="bib27" id="ref24">27</reflink>],[<reflink idref="bib29" id="ref25">29</reflink>]–[<reflink idref="bib31" id="ref26">31</reflink>]. Nonetheless, the thorough involvement of Cx43 in the pathogenesis and development of breast cancer is far from being elucidated, hence more efforts are needed to comprehensively approach this topic.</p> <p>BreastAnalyser is intended to fulfill the following requirements: (<reflink idref="bib1" id="ref27">1</reflink>) provide a friendly GUI to interactively work with the images; (<reflink idref="bib2" id="ref28">2</reflink>) use image analysis and machine learning algorithms to automatically recognise DAB-brown-stained proteins by immunohistochemistry, and detect and classify the nuclei in an image or ROIs; (<reflink idref="bib3" id="ref29">3</reflink>) automatically estimate various statistical measures and counts in the IHC images; (<reflink idref="bib4" id="ref30">4</reflink>) allow data sharing among researchers and review the results at any time; and (<reflink idref="bib5" id="ref31">5</reflink>) be fast enough to analyse images in real time. In relation to the previously mentioned softwares, BreastAnalyser represents a more straigthforward alternative, where specific post-processing concerns can be manually addressed straightway. BreastAnalyser, differently from QuPath, does not require modification of the annotations in order to correct mistakes regarding DAB-brown signal or nuclei detection/classification. Instead, BreastAnalyser allows an immediate adjustment of brown signal-pixels, using the add/delete/draw tools, and of located cell nuclei, by means of the add/delete/modify classification tools, with just a click of a button.</p> <p>Graph: Figure 2Representative Cx43-IHC images and hematoxylin nuclear counter-stain (purple) of invasive luminal B HER2- breast cancer tumours tested for Oncotype DX and categorized in low and high recurrence risk categories.</p> <hd id="AN0175756719-3">Biological materials and experimental methodology</hd> <p>Biological samples used for Cx43 assessment comprise paraffin-embedded invasive breast tumours and normal mammary tissue controls, and are part of the A Coruña Biobank and our private collection of human biological samples. The samples used in our biomedical studies belong to white breast cancer female patients from the A Coruña sanitary area (NW Spain), with a median age of 52 years, being 48 % premenopausic and 52 % postmenopausic. Additionally, a variety of breast tumour immunohistochemistries were obtained from the collection of human biological samples of the Albacete General University Hospital in order to evaluate software inter-laboratory operation. Antigens detected are relevant in breast cancer anatomo-histopathological analysis and encompass RING1, RING4, CD99, CD31, CD177, EMA and cytokeratins 1 and 7. Immunohistochemistry was performed according to the standard protocols, described in detail in Section 1 of the supplementary material.</p> <p>The IHC samples of INIBIC lab were photographed under an Olympus BX61 microscope coupled to an Olympus DP71 digital camera using a magnification of 40X, resulting in images of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>2040</mn><mo>×</mo><mn>1436</mn></mrow></math> </ephtml> pixels. For the CRIB lab, images were acquired by a Nikon Eclipse 80i microscope coupled to a DXM1200C digital camera (Nikon), using magnification of 20<emph>X</emph> and 40<emph>X</emph>, resulting in images of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>1372</mn><mo>×</mo><mn>1024</mn></mrow></math> </ephtml> pixels. One image was analysed from each sample-patient for the biomedical results discussion ("Biomedical results" section). The ROI were selected and manually drawn by the experts according to their anatomo-histopathological relevance (i.e.tumour cell nests or cords, healthy duct-lobular units and acini, etc).</p> <p>We distributed the samples into three groups in order to perform different types of studies. The group <bold>SET 1</bold> is composed of 33 breast cancer patient samples derived from core needle biopsies of invasive tumours (grades 2–3) not subjected to chemotherapy or radiotherapy, in order to determine potential variations in Cx43 expression depending on the breast cancer subtype: luminal A (6 samples), luminal B HER2+ (5 samples), luminal B HER2- (8 samples), non-luminal HER2+ (7 samples) and triple negative (7 samples). This set also comprises 6 samples of normal healthy breast tissue for control. Specifically, <bold>SET 1</bold> includes 47 images extracted from the 39 samples, with Cx43 signal in brown, ranging from 0.06% to 38.3%. According to the expression and localization of Cx43 in human breast cancer (BC) tissues, it is progressively downregulated with increasing subtype malignancy. Figure 1 shows representative images of Cx43-IHC (brown) for each subtype, as well as for normal mammary tissue, where Cx43 immunostaining presents the highest expression levels, with membranous and cytoplasmic staining in both the myoepithelial and luminal layers of mammary ducts (1st panel) and acini (2nd panel). Luminal A breast cancer, the least aggressive subtype, is characterized by high Cx43 expression, mainly cytoplasmic (3rd panel), followed closely by luminal B HER2+ samples (4th panel). Luminal B HER2- tumours show lower Cx43 levels (cytoplasmic) (5th panel), whereas non luminal HER2+ (6th panel) and triple negative samples (last panel) score the lowest Cx43 levels, with almost no protein expression in the latter. <bold>SET 2</bold> comprises invasive breast cancer luminal B HER2- samples, grades 2–3, assessed for Oncotype DX Breast Recurrence Score Test: 6 scored high risk (scores 31–46) and 6 scored low risk (scores 9–20) for cancer recurrence. In this case, the purpose is to discern possible differences in Cx43 levels between high and low risk-scoring samples according to this test. Figure 2 shows representative Cx43-IHC images. It encompasses 32 images extracted from the 12 samples. <bold>SET 3</bold> contains 14 images extracted from 14 samples provided by the CRIB lab, which were purposely included for inter-laboratory software performance comparison. Figure 3 shows representative images of relevant antigens routinely assessed in breast cancer anatomopathological analysis.</p> <p>Graph: Figure 3Representative IHC images from the CRIB lab.</p> <hd id="AN0175756719-4">Computing methods</hd> <p>BreastAnalyser is a desktop application that runs on general purpose computers under Linux or Windows operating systems. It has been written in the C++ programming language using the GTK+ (GIMP Tool Kit) library (https://<ulink href="http://www.gtk.org/">www.gtk.org/</ulink>) to develop the GUI and the OpenCV library (https://opencv.org/) to generate the automatic image processing algorithms. Figure 4 shows its GUI with a typical IHC image loaded, processed and reviewed by the expert, and with the lateral panel displayed.</p> <p>Graph: Figure 4Screenshot of the BreastAnalyser GUI. In the region of analysis (defined by the cyan lines), the colour of the dots shows the category of the nuclei: yellow (highly stained), pink (moderately stained), blue (low stained) and green (not stained). The brown areas are surrounded by black lines, and the white regions are unstained zones inside the brown areas.</p> <p>The "Functionality of BreastAnalyser" section describes the architecture and the main functionality provided by BreastAnalyser. In "Algorithms to recognise objects of interest" section details the image analysis algorithms to recognise the objects of interest (DAB-brown regions and nuclei) and "Machine learning methods" section describes the classification of the staining level of the nuclei.</p> <p>Graph: Figure 5Flowchart containing the main tasks of BreastAnalyser.</p> <hd id="AN0175756719-5">Functionality of BreastAnalyser</hd> <p>BreastAnalyser is a modular and extensible software composed by a GUI layer with editing tools to interact with the user; a logic layer that contains modules to automatically process the image and to calculate the statistical results; and a persistence layer to store all the data needed and calculated by the software. The software includes modules to: (<reflink idref="bib1" id="ref32">1</reflink>) store the image overlays, that contain the analysis supervised by the experts, in XML (Extensible Markup Language) files; and (<reflink idref="bib2" id="ref33">2</reflink>) save the statistical results, calculated from the overlays, in CSV (Comma-Separated Values) files.</p> <p>Figure 5 shows a flowchart with the main functionality of BreastAnalyser. A typical working session for an user should have the following actions: (<reflink idref="bib1" id="ref34">1</reflink>) open an image; (<reflink idref="bib2" id="ref35">2</reflink>) automatically detect the DAB-brown regions; (<reflink idref="bib3" id="ref36">3</reflink>) automatically detect the nuclei in the image or into ROIs manually drawn and labeled by the expert as <emph>TUMOR</emph>, <emph>NON-TUMOR</emph> or <emph>STROMA</emph>; (<reflink idref="bib4" id="ref37">4</reflink>) automatically classify the detected or marked nuclei on the image; (<reflink idref="bib5" id="ref38">5</reflink>) go to expert's supervision, as described below; (<reflink idref="bib6" id="ref39">6</reflink>) save the overlays drawn on the image into XML files; (<reflink idref="bib7" id="ref40">7</reflink>) export the statistical measures and counts to CSV files; and (<reflink idref="bib8" id="ref41">8</reflink>) at any time the user can set preferences, set calibration and diameters, save joined results of a set of images, or train the classifier.</p> <p>Once the image is loaded, the buttons of the lateral panel can be used to automatically process the image following the instructions included in the user guide, provided as supplementary material. The detection and classification of the nuclei can be performed on the whole image or into ROIs drawn and labeled by the experts, depending on the objectives of the study. The classifier can be run in order to automatically provide a category for each detected nuclei. The available categories for the nuclei are "without staining" or "with low/medium/high staining". Due to the inherent complexity of these images, the automatic processing may not be optimal for the expert, so BreastAnalyser provides an easy way to review the recognised objects in the images through the following editing tools (semi-automation): (<reflink idref="bib1" id="ref42">1</reflink>) delete a set of selected objects; (<reflink idref="bib2" id="ref43">2</reflink>) change the category of selected objects; and (<reflink idref="bib3" id="ref44">3</reflink>) add new nuclei specifying their category labels. The overlays set on the images contain their analysis information and they must be saved into the XML file in order to use other software functionalities, such as review the analysis, export joined results or train the classifier.</p> <p>The working preferences of BreastAnalyser can be set going to the menu <emph>File</emph><ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">→</mo></math> </ephtml><emph>Set preferences</emph>, which allows to determine: (<reflink idref="bib1" id="ref45">1</reflink>) the working directories for images, overlays and results; (<reflink idref="bib2" id="ref46">2</reflink>) the width of points and lines; and (<reflink idref="bib3" id="ref47">3</reflink>) the colour of the overlays for each category. By default the measures on the images are provided in pixels. In order to obtain the results in real values, the user must set the spatial calibration, which is the relation between pixels in the image and real values (micrometers). The user must provide the minimum and maximum diameter of nuclei to be detected for an optimal operation of the automatic algorithms. These diameters can be set by writing in the <emph>Preferences</emph> dialog or graphically by drawing a straight line with the editing tools of the lateral panel. The preferences can be permanently stored for future working sessions. BreastAnalyser allows to export joined results of a set of images going to the menu <emph>Analysis</emph><ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">→</mo></math> </ephtml><emph>XML Files</emph>, which opens a dialog screen to select the XML files and the output CSV file. To do this task, the images have to be analysed and supervised by the expert, and finally the overlays must be saved in XML files, one per image. BreastAnalyser also allows to train the classifier going to the menu <emph>Analysis</emph><ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">→</mo></math> </ephtml><emph>Train classifier</emph>. More details of the BreastAnalyser use can be read in the user guide.</p> <hd id="AN0175756719-6">Algorithms to recognise objects of interest</hd> <p>The recognition of DAB-brown or detection of nuclei on the image are segmentation problems where the DAB-brown regions or nuclei are the objects and the remaining area is the background. Image segmentation is an important topic in computer vision[<reflink idref="bib32" id="ref48">32</reflink>], which is based on the analysis of the properties of homogeneity and discontinuity within the pixels of the image. Some image properties frequently used are colour, grey level or texture. The homogeneity paradigm develops the region detection algorithms, which try to keep the properties of the image constant within the regions. The discontinuity paradigm develops the edge-based algorithms, which attempt to find the position of the discontinuities in the image properties between the objects and the background. In this subsection, we describe the image segmentation algorithms included in BreastAnalyser. In "RBA to detect the nuclei" section describes the region-based BrownDetector algorithm to recognise the DAB-brown regions in the image. In BrownDetector algorithm to recognise DAB-brown signal" and "EBA algorithm to detect the nuclei" sections describe the region-based (RBA) algorithm and edge-based (EBA) algorithms used to detect nuclei. In the case of nuclei segmentation, experts only want to count the number of nuclei. So, the nuclei recognitions are transformed to points in BreastAnalyser. One of the main challenges to automatically process the pathology images is the colour variation among images due to differences in preparation and digitalization of samples, which can influence the performance of image analysis algorithms[<reflink idref="bib33" id="ref49">33</reflink>],[<reflink idref="bib34" id="ref50">34</reflink>]. Normalization is often applied to original images in order to standardize them for processing[<reflink idref="bib35" id="ref51">35</reflink>], but this pre-processing requires time and it might not be suitable for interactive applications. In this paper, we propose algorithms that are parametrized by the properties of each image. In this way the normalization is implicit in the algorithm itself, avoiding the normalization time.</p> <p>Graph: Figure 6Examples of the automatic processing of IHC images using the BrownDetector algorithm for different types of processing: I is the original image, IS=IaIb being the Ia and Ib the a and b channel of Lab original image; and IB is the thresholded image (see the text for a detailed description). The IOut image shows the set of contours R (in green) and H (in red) recognised overlapped to the original image.</p> <p>Graph: Algorithm 1BrownDetector to recognise the DAB-brown regions.</p> <hd id="AN0175756719-7">BrownDetector algorithm to recognise DAB-brown signal</hd> <p>The proposed algorithm, called BrownDetector, to recognise the DAB-brown regions falls into the region based paradigm and it is a combination of computer vision techniques parameterized using the information of each image. The main steps of BrownDetector algorithm are summarized in the algorithm 1. Let <emph>I</emph>(<emph>x</emph>, <emph>y</emph>), with <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>x</mi><mo>=</mo><mn>1</mn><mo>,</mo><mo>...</mo><mo>,</mo><mi>N</mi></mrow></math> </ephtml> , and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>y</mi><mo>=</mo><mn>1</mn><mo>,</mo><mo>...</mo><mo>,</mo><mi>M</mi></mrow></math> </ephtml> , be the original RGB image and let <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mi>P</mi><mo>=</mo><mo stretchy="false">{</mo><mi>L</mi><mi>e</mi><mi>s</mi><mi>s</mi><mo>,</mo><mi>M</mi><mi>i</mi><mi>d</mi><mi>d</mi><mi>l</mi><mi>e</mi><mo>,</mo><mi>M</mi><mi>o</mi><mi>r</mi><mi>e</mi><mo stretchy="false">}</mo></mrow></math> </ephtml> be the type of processing chosen by the expert, as it can be seen in the lateral panel of the GUI. Let <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">R</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msub><mi>R</mi><mi>i</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>R</mi></msub></msubsup></mrow></math> </ephtml> be the set of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>N</mi><mi>R</mi></msub></math> </ephtml> DAB-brown regions automatically recognised, and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">H</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msub><mi>H</mi><mi>j</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>H</mi></msub></msubsup></mrow></math> </ephtml> be the set of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>N</mi><mi>H</mi></msub></math> </ephtml> regions, which represent holes inside the DAB regions. Firstly, the original RGB image <emph>I</emph> is transformed into the Lab colour space, that is more robust to illuminance variance[<reflink idref="bib35" id="ref52">35</reflink>]. Hence, the Lab colour space is intended to be perceptually uniform, i. e. changes in the numerical values are similar to the perceived change in colour. The <emph>L</emph> channel is associated to the lightness and the <emph>a</emph> and <emph>b</emph> channels are associated with the chrominance, specifically with the redness and yellowness respectively. In painting, the combination of red and yellow provides different shades of brown. So, the channels <emph>a</emph> and <emph>b</emph> are multiplied to build an image with different shades of brown developing the float image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">ab</mi></mrow></msub></math> </ephtml> , which is transformed into the grey level image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> doing contrast stretching. In this image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> , the lower values (darker pixels) correspond to the brownest pixels in the original image (see second column in Fig. 6 for visual examples). The <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> image is thresholded in order to segment the foreground (DAB-brown regions) from the background. But, the selection of the optimal threshold value is a challenging task and many times it is determined by trial and error in the literature. In our approach, this threshold <emph>T</emph> is determined for each image from its statistical characteristics. Let <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>H</mi><mi>S</mi></msub></math> </ephtml> be the histogram of image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> , we define the threshold <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>T</mi><mi>h</mi></msub></math> </ephtml> as the lowest grey level, starting from the 0, in the cumulative histogram in which the 1% of the total number of pixels in the image is achieved. Other thresholds are calculated using the multi-level method proposed by Otsu[<reflink idref="bib36" id="ref53">36</reflink>], which selects a <emph>T</emph> that minimizes intra-class intensity variance maximizing inter-class variance. Specifically, let <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>1</mn></msub></math> </ephtml> , <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>2</mn></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>3</mn></msub></math> </ephtml> be the thresholds calculated after applying the multi-level Otsu's method with three thresholds to <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> (function Otsu3Thres in algorithm 1). The optimal threshold <emph>T</emph> to binarize the image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> depends on the type of processing <emph>TP</emph> and it is calculated as <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><msub><mi>T</mi><mi>h</mi></msub></mrow></math> </ephtml> when <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mi>P</mi><mo>=</mo><mi>L</mi><mi>e</mi><mi>s</mi><mi>s</mi></mrow></math> </ephtml> ; if <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mi>P</mi><mo>=</mo><mi>M</mi><mi>i</mi><mi>d</mi><mi>d</mi><mi>l</mi><mi>e</mi></mrow></math> </ephtml> , then <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><msub><mi>t</mi><mn>1</mn></msub><mo>-</mo><mrow><mo stretchy="false">(</mo><msub><mi>t</mi><mn>1</mn></msub><mo>-</mo><msub><mi>T</mi><mi>h</mi></msub><mo stretchy="false">)</mo></mrow><mo stretchy="false">/</mo><mn>2</mn></mrow></math> </ephtml> ; and finally, if <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mi>P</mi><mo>=</mo><mi>M</mi><mi>o</mi><mi>r</mi><mi>e</mi></mrow></math> </ephtml> , then <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><msub><mi>t</mi><mn>2</mn></msub><mo>-</mo><mrow><mo stretchy="false">(</mo><msub><mi>t</mi><mn>2</mn></msub><mo>-</mo><msub><mi>t</mi><mn>1</mn></msub><mo stretchy="false">)</mo></mrow><mo stretchy="false">/</mo><mn>2</mn></mrow></math> </ephtml> . Normally, the best processing is achieved with the option <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mi>P</mi><mo>=</mo><mi>M</mi><mi>i</mi><mi>d</mi><mi>d</mi><mi>l</mi><mi>e</mi></mrow></math> </ephtml> , but the remaining options provide adequate responses to extreme cases, such as images with very low or great positivity. The image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>S</mi></msub></math> </ephtml> is transformed into a binary image using the threshold <emph>T</emph>, i.e. if <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>I</mi><mi>S</mi></msub><mrow><mo stretchy="false">(</mo><mi>x</mi><mo>,</mo><mi>y</mi><mo stretchy="false">)</mo></mrow><mo><</mo><mi>T</mi></mrow></math> </ephtml> the output is 255 and 0 otherwise. This process is called inverse thresholding. Mathematical morphology is commonly used for morphological processing of images, which is composed by the basic filters: dilation, erosion, open and close. Initially, it was defined to binary images in order to fill/remove objects smaller than the size of structural element[<reflink idref="bib32" id="ref54">32</reflink>]. So, we use a close filter with masksize 5 to fill small holes in the thresholded images developing the image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>B</mi></msub></math> </ephtml> . The contours of the DAB-brown regions (set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msubsup><mrow><mo stretchy="false">{</mo><msub><mi>R</mi><mi>i</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>R</mi></msub></msubsup></math> </ephtml> ) and holes into DAB-brown regions (set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msubsup><mrow><mo stretchy="false">{</mo><msub><mi>H</mi><mi>j</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>H</mi></msub></msubsup></math> </ephtml> ) are extracted from <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>B</mi></msub></math> </ephtml> using the algorithm proposed by Suzuki and Be[<reflink idref="bib37" id="ref55">37</reflink>]. The number of DAB regions is the dimension of the set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi mathvariant="script">R</mi></math> </ephtml> and the DAB-brown area percentage is calculated adding up the area of regions in set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi mathvariant="script">R</mi></math> </ephtml> and subtracting the area of regions in <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi mathvariant="script">H</mi></math> </ephtml> . Figure 6 shows visual examples of the performance of BrownDetector algorithm in different breast cancer samples using green colour for DAB regions and red for holes inside DAB regions: (<reflink idref="bib1" id="ref56">1</reflink>) Luminal B HER2+ (top row): using the option <emph>Middle</emph>, the threshold <emph>T</emph> calculated is 54 and the percentage of DAB area is 2.19%; (<reflink idref="bib2" id="ref57">2</reflink>) healthy breast (second row) using option <emph>More</emph>, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><mn>93</mn></mrow></math> </ephtml> and the 27.4% area occupied by DAB regions; (<reflink idref="bib3" id="ref58">3</reflink>) triple negative (third row) using the option <emph>Middle</emph>, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><mn>62</mn></mrow></math> </ephtml> and 5.59% of DAB area; and (<reflink idref="bib4" id="ref59">4</reflink>) HER2+ (bottom row), using option <emph>Less</emph>, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><mo>=</mo><mn>69</mn></mrow></math> </ephtml> and 0.93% of DAB area.</p> <hd id="AN0175756719-8">RBA to detect the nuclei</hd> <p>Graph: Algorithm 2RBA algorithm to detect nuclei.</p> <p>Graph: Figure 7Examples of the automatic processing of immunohistochemical images using RBA to detect the nuclei (see algorithm 2 for the meaning of IG , IB1 , Iin and IB3). The image IOut shows the set of nuclei positions P overlapped to the original image (in black and green the nuclei without staining and low staining, respectively, using the classifier).</p> <p>The RBA algorithm attempts to segment the nuclei using the region-based segmentation paradigm. Afterwards, the mass center of each segmented region is considered as the detection of each nucleus. The main steps of RBA are summarised in algorithm 2. Let <emph>I</emph> be the original RGB immunohistochemical images and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></math> </ephtml> the minimum and maximum diameter of the nuclei to be detected. Firstly, the <emph>I</emph> image is transformed to the Lab colour space and the colour <emph>a</emph> and <emph>b</emph> channels, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>a</mi></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>b</mi></msub></math> </ephtml> respectively, are considered, which are processed by a median filter with masksize 3 in order to attenuate random noise. Secondly, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> is built by adding the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>a</mi></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>b</mi></msub></math> </ephtml> images and normalizing the result to the range [0,255] (see a visual example of the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> image in Fig. 7). The nuclei appear as bright spots inside the background, but the brightness of the nuclei outside the DAB-brown regions is higher than that of the inner ones. So, we detect the nuclei using two stages: (<reflink idref="bib1" id="ref60">1</reflink>) threshold <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> image using <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>3</mn></msub></math> </ephtml> to detect the nuclei outside of the DAB-brown regions; and (<reflink idref="bib2" id="ref61">2</reflink>) create a mask image, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">mask</mi></mrow></msub></math> </ephtml> , containing only the DAB regions and apply thresholding to detect the nuclei inside the DAB regions. In all cases, the threshold values are determined automatically using the multi-level method of Otsu[<reflink idref="bib36" id="ref62">36</reflink>], function Otsu3Thres and Otsu3ThresWB in algorithm 2. In the first step, the optimal threshold <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mi>G</mi></msub></math> </ephtml> to segment the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> images is the third threshold ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>3</mn></msub></math> </ephtml> in algorithm 2, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>t</mi><mn>3</mn></msub><mo>=</mo><mn>132</mn></mrow></math> </ephtml> in Fig. 7). The resulting image after thresholding is processed by a morphological open filter using a masksize <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>m</mi><mo>=</mo><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></mrow></math> </ephtml> in order to remove objects smaller than the structural element and split touching nuclei, developing the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi>B</mi><mn>1</mn></mrow></msub></math> </ephtml> image (see Fig. 7). The set of nuclei positions, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi mathvariant="script">B</mi></math> </ephtml> , includes the mass centers of the white regions in <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi>B</mi><mn>1</mn></mrow></msub></math> </ephtml> with diameter <emph>d</emph> such that <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub><mo><</mo><mi>d</mi><mo><</mo><msub><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></mrow></math> </ephtml> .</p> <p>In the second step, the value <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>t</mi><mn>2</mn></msub></math> </ephtml> (line 6 of algorithm 2) is used to inverse threshold <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> developing the image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">bin</mi></mrow></msub></math> </ephtml> (for example in Fig. 7 we use <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>t</mi><mn>2</mn></msub><mo>=</mo><mn>88</mn></mrow></math> </ephtml> ). The <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">bin</mi></mrow></msub></math> </ephtml> image is post-processed by a close filter of size <emph>m</emph> in order to fill holes smaller than the nuclei size. We only keep the external regions larger than an area <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>a</mi><mo>=</mo><mn>2</mn><mi>π</mi><msubsup><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow><mn>2</mn></msubsup></mrow></math> </ephtml> to create a mask image, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">mask</mi></mrow></msub></math> </ephtml> , which multiplied by the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mi>G</mi></msub></math> </ephtml> image develops the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">In</mi></mrow></msub></math> </ephtml> image in Fig. 7. In these images the DAB-brown regions are seen, while the remaining dark areas in image are associated to background. We apply again the multi-level Otsu method to the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">In</mi></mrow></msub></math> </ephtml> image without considering the black pixels (function Otsu3ThresWB in line 14 of algorithm 2) and use the third value <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msubsup><mi>t</mi><mn>3</mn><mo>′</mo></msubsup></math> </ephtml> to threshold the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">In</mi></mrow></msub></math> </ephtml> image ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msubsup><mi>t</mi><mn>3</mn><mo>′</mo></msubsup><mo>=</mo><mn>89</mn></mrow></math> </ephtml> for example in Fig. 7). The <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi>B</mi><mn>3</mn></mrow></msub></math> </ephtml> image in Fig. 7 shows the resulting image after applying an open filter with masksize <emph>m</emph> to remove noisy regions. The set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi mathvariant="script">C</mi></math> </ephtml> of nuclei positions is obtained by applying the size filter to the white regions in image <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi>B</mi><mn>3</mn></mrow></msub></math> </ephtml> (i.e. regions whose diameter <emph>d</emph> satisfies <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub><mo><</mo><mi>d</mi><mo><</mo><msub><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></mrow></math> </ephtml> , line 16 of algorithm 2). Finally, we apply an overlapping test to remove the nuclei which have been detected by the two stages. So, <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">P</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>P</mi></msub></msubsup></mrow></math> </ephtml> will be the nuclei that are farther from each other than <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> , i.e. <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mo stretchy="false">{</mo><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub><mo>∈</mo><mi mathvariant="script">B</mi><mo>,</mo><msub><mi mathvariant="bold">c</mi><mi>j</mi></msub><mo>∈</mo><mi mathvariant="script">C</mi><mo stretchy="false">|</mo></mrow></math> </ephtml> distance( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub></math> </ephtml> , <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="bold">c</mi><mi>j</mi></msub></math> </ephtml> ) <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mo>></mo><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub><mrow><mo stretchy="false">}</mo></mrow></mrow></math> </ephtml> . The <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>I</mi><mrow><mi mathvariant="italic">out</mi></mrow></msub></math> </ephtml> image in Fig. 7 shows the original image with the nuclei positions, set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">P</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>P</mi></msub></msubsup></mrow></math> </ephtml> , overlapped as circles for visualization purpose. The colour means the staining level of the nuclei (black and green for non-stained and low stained nuclei, respectively).</p> <p>Graph: Algorithm 3EBA algorithm to detect nuclei.</p> <hd id="AN0175756719-9">EBA algorithm to detect the nuclei</hd> <p>The EBA algorithm to detect nuclei is based on the edge-based paradigm. Specifically, EBA is a modified version of the Canny filter, proposed by Canny[<reflink idref="bib38" id="ref63">38</reflink>], already used in our software Govocitos[<reflink idref="bib39" id="ref64">39</reflink>]. Algorithm 3 summarizes the EBA algorithm, that only uses one Canny filter tuned with a Gaussian smoothing width <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>σ</mi><mo>=</mo><mn>4</mn></mrow></math> </ephtml> in order to remove image noise. The thresholds of the hysteresis process in the Canny filter are automatically calculated from the image characteristics using rates of 0.4 and 0.6 for the lower and higher thresholds, respectively. The output of the Canny filter is a set of edges <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi mathvariant="script">E</mi><mi>C</mi></msub><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msup><mi>E</mi><mi>i</mi></msup><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>C</mi></msub></msubsup></mrow></math> </ephtml> , that is post-processed as follows to calculate the positions of nuclei. The set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">E</mi><mi>C</mi></msub></math> </ephtml> is filtered using the minimum ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> ) and maximum ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></math> </ephtml> ) diameters for the nucleus provided by the expert. If we considered that the nucleus is rounded, its area can be approximated by a circle with area <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>A</mi><mo>=</mo><mi>π</mi><msup><mi>d</mi><mn>2</mn></msup><mo stretchy="false">/</mo><mn>4</mn></mrow></math> </ephtml> , being <emph>d</emph> the diameter of the nucleus edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msup><mi>E</mi><mi>i</mi></msup></math> </ephtml> . So, we use the following criteria: (<reflink idref="bib1" id="ref65">1</reflink>) the area <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>A</mi><mo stretchy="false">(</mo><msup><mi>E</mi><mi>i</mi></msup><mo stretchy="false">)</mo></mrow></math> </ephtml> enclosed by <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msup><mi>E</mi><mi>i</mi></msup><mo>∈</mo><msub><mi mathvariant="script">E</mi><mi>C</mi></msub></mrow></math> </ephtml> satisfies that <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>A</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub><mo><</mo><mi>A</mi><mrow><mo stretchy="false">(</mo><msup><mi>E</mi><mi>i</mi></msup><mo stretchy="false">)</mo></mrow><mo><</mo><msub><mi>A</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></mrow></math> </ephtml> , being <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>A</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>A</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></math> </ephtml> the areas calculated using <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">max</mi></mrow></msub></math> </ephtml> , respectively; and (<reflink idref="bib2" id="ref66">2</reflink>) the roundness of edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msup><mi>E</mi><mi>i</mi></msup></math> </ephtml> , (round function in algorithm 3) is lower than 1.4, since the circle roundness is 1. This post-processing filtering creates a new set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi mathvariant="script">E</mi><mi>F</mi></msub><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msup><mi>E</mi><mi>i</mi></msup><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>F</mi></msub></msubsup></mrow></math> </ephtml> with <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>N</mi><mi>F</mi></msub><mo>≤</mo><msub><mi>N</mi><mi>C</mi></msub></mrow></math> </ephtml> . To avoid two detections of some nuclei, an overlapping test is applied: an edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msup><mi>E</mi><mi>i</mi></msup><mo>∈</mo><msub><mi mathvariant="script">E</mi><mi>F</mi></msub></mrow></math> </ephtml> is removed if there is another edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msup><mi>E</mi><mi>i</mi></msup></math> </ephtml> with <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="monospace">distance</mi><mrow><mo stretchy="false">(</mo><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub><mo>,</mo><msub><mi mathvariant="bold">c</mi><mi>j</mi></msub><mo stretchy="false">)</mo></mrow><mo>≤</mo><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></mrow></math> </ephtml> . After this overlapping test, a new set <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi mathvariant="script">E</mi><mi>O</mi></msub><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msup><mi>E</mi><mi>i</mi></msup><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>O</mi></msub></msubsup></mrow></math> </ephtml> with <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msub><mi>N</mi><mi>O</mi></msub><mo>≤</mo><msub><mi>N</mi><mi>F</mi></msub></mrow></math> </ephtml> is created. Finally, we use a pre-trained classifier to predict if an edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msup><mi>E</mi><mi>i</mi></msup><mo>∈</mo><msub><mi mathvariant="script">E</mi><mi>O</mi></msub></mrow></math> </ephtml> is a true nucleus or a false positive (see the "Machine learning methods" section for details). The input of the classifier is a numerical vector including the mean values of the Lab colour space channel in a square of size <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>d</mi><mrow><mi mathvariant="italic">min</mi></mrow></msub></math> </ephtml> centered in the centroid <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub></math> </ephtml> of edge <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><msup><mi>E</mi><mi>i</mi></msup><mo>∈</mo><msub><mi mathvariant="script">E</mi><mi>O</mi></msub></mrow></math> </ephtml> . If the classification prediction is a nucleus, the centroid is added to the set of nuclei <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">P</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msub><mi mathvariant="bold">c</mi><mi>i</mi></msub><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><msub><mi>N</mi><mi>P</mi></msub></msubsup></mrow></math> </ephtml> . In order to process an user-defined ROI, the EBA algorithm is applied only on a rectangle enclosing the ROI. Examples of the visual performance of the EBA algorithm can be seen in the user guide.</p> <hd id="AN0175756719-10">Machine learning methods</hd> <p>A classifier is a machine learning method for the automatic prediction of discrete values (output categories) based on data examples. The classifier learns to predict the outcome category as a function of the input data in a process called "training", that uses a collection of examples, each composed by the input data and the outcome value. During training, the model changes the values of its parameters in order to predict an outcome near to the true value for the training data, i.e., to give a reliable prediction for these data. The trained model is expected to generalise its predictions with reliability to new input data not used during training.</p> <p>BreastAnalyser includes two classifiers: ClassifierFP, used to discriminate between valid (true-positive) and non-valid (false positive) nuclei; and ClassifierNucleus, used to discriminate among different staining levels in the nuclei: with high, medium or low staining and not stained. Both classifiers use the support vector machine (SVM) with radial basis function (RBF) kernel, selected because it is one of the best-performing machine learning models for classification[<reflink idref="bib40" id="ref67">40</reflink>]. Specifically, BreastAnalyser uses the LibSVM implementation[<reflink idref="bib41" id="ref68">41</reflink>], accessed through its C++ binding. The SVM used by ClassifierNucleus for the staining level of the nuclei can be re-trained from the GUI at any time through the submenu <emph>Classification</emph><ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">→</mo></math> </ephtml><emph>Train classifier</emph>. In this case, the SVM is trained using a collection of nuclei randomly selected from the XML files provided by the user. In order to avoid an excessively slow training, a maximum number of 1,000 nuclei are selected. Whenever posible, the nuclei are selected in similar numbers, requiring a minimum number of 10 nuclei, for each staining category. BreastAnalyser performs the tuning of the two hyper-parameters of the SVM (regularization <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi>λ</mi></math> </ephtml> and RBF kernel spread <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi>σ</mi></math> </ephtml> ) using the grid-search method. The performance is evaluated by the Cohen kappa statistic[<reflink idref="bib42" id="ref69">42</reflink>], which measures the coincidence between the true and predicted category excluding the agreement by chance. Kappa (in %) is defined as:</p> <olist> <item> <ephtml> <math display="block" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mtable><mtr><mtd columnalign="right"><mstyle displaystyle="true" scriptlevel="0"><mrow><mi>k</mi><mi>a</mi><mi>p</mi><mi>p</mi><mi>a</mi><mo>=</mo><mn>100</mn><mfrac><mrow><msub><mi>p</mi><mi>a</mi></msub><mo>-</mo><msub><mi>p</mi><mi>e</mi></msub></mrow><mrow><mi>s</mi><mo>-</mo><msub><mi>p</mi><mi>e</mi></msub></mrow></mfrac><mo>,</mo><mspace width="2em" /><msub><mi>p</mi><mi>a</mi></msub><mo>=</mo><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><msub><mi>N</mi><mrow><mi mathvariant="italic">ii</mi></mrow></msub><mo>,</mo><mspace width="2em" /><msub><mi>p</mi><mi>e</mi></msub><mo>=</mo><mfrac><mn>1</mn><msup><mi>N</mi><mn>2</mn></msup></mfrac><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><mfenced close=")" open="("><munderover><mo>∑</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><msub><mi>N</mi><mrow><mi mathvariant="italic">ij</mi></mrow></msub></mfenced><mfenced close=")" open="("><munderover><mo>∑</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><msub><mi>N</mi><mrow><mi mathvariant="italic">ji</mi></mrow></msub></mfenced><mo>,</mo><mspace width="1em" /><mi>s</mi><mo>=</mo><munderover><mo>∑</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><munderover><mo>∑</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>C</mi></munderover><msub><mi>N</mi><mrow><mi mathvariant="italic">ij</mi></mrow></msub></mrow></mstyle></mtd></mtr></mtable></mrow></math> </ephtml> </item> </olist> <p>Graph</p> <p>where <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>N</mi><mrow><mi mathvariant="italic">ij</mi></mrow></msub></math> </ephtml> is the number of nuclei of category <emph>i</emph> that are assigned by the SVM to category <emph>j</emph>, while <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>C</mi><mo>=</mo><mn>4</mn></mrow></math> </ephtml> is the number of categories and <emph>N</emph> is the number of nuclei. The values of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi>λ</mi></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mi>σ</mi></math> </ephtml> used for hyper-parameter tuning are: <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>λ</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msup><mn>2</mn><mrow><mn>2</mn><mi>i</mi><mo>-</mo><mn>7</mn></mrow></msup><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mn>10</mn></msubsup></mrow></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>σ</mi><mo>=</mo><msubsup><mrow><mo stretchy="false">{</mo><msup><mn>2</mn><mrow><mo>-</mo><mo stretchy="false">(</mo><mi>i</mi><mo>+</mo><mn>1</mn><mo stretchy="false">)</mo><mo stretchy="false">/</mo><mn>2</mn></mrow></msup><mo stretchy="false">}</mo></mrow><mrow><mi>i</mi><mo>=</mo><mo>-</mo><mn>15</mn></mrow><mn>0</mn></msubsup></mrow></math> </ephtml> . For each combination of hyper-parameter values, the SVM is trained using the <emph>K</emph>-fold cross-validation methodology with <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>K</mi><mo>=</mo><mn>4</mn></mrow></math> </ephtml> , so that <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>K</mi><mo>-</mo><mn>1</mn><mo>=</mo><mn>3</mn></mrow></math> </ephtml> folds are used to train the SVM, and the remaining fold is used to calculate the kappa of the trained SVM. The training and prediction are performed <emph>K</emph> times, rotating the folds each time (i.e., in the first trial folds 1-3 are used for training and fold 4 for test; the second trial uses folds 2-4 to train and fold 1 for test; and so on) and averaging kappa over the <emph>K</emph> test folds. The process is repeated for all the combinations of hyper-parameter values, and the one that achieves the highest average kappa is selected. Finally, the SVM is trained over the whole collection of nuclei, using the selected combination of hyper-parameter values, and then it is ready to predict the category for new nuclei.</p> <p>The default ClassifierNucleus included in BreastAnalyser was trained with a selected set of samples provided by INIBIC lab composed by 26 IHC images (from patients not included in <bold>SET 1</bold> nor <bold>SET 2</bold>) containing 1,359 nuclei with different staining levels: 160 high, 195 medium, 328 low and 676 nuclei without staining, selected attempting to represent all the posible variability in the staining level of the nuclei. The training followed the methodology described in the previous paragraph.</p> <p>The feature vector used by both SVM classifiers contains the mean value of each channel in the Lab image over a neighborhood of the nucleus. This neighborhood is centered on the centroid of the nucleus and its size is equal to the minimum diameter of nuclei, fixed by the expert using the BreastAnalyser GUI.</p> <hd id="AN0175756719-11">Results</hd> <p>BreastAnalyser has been used since 2021 in the daily research work of experts at INIBIC lab in order to evaluate the software in a real environment. Since 2022, this software is also being used in CRIB lab in order to assess inter-laboratory operation. The interaction of clinical staff with BreastAnalyser was logged into the XML files to perform a statistical evaluation of the automatic algorithms of image processing and machine learning included in the software. The main goal of this research is the evaluation of the robustness and versatility of BreastAnalyser software when exposed to highly variable IHC images, representative of the complexity and inter-patient variability of a real-world clinical scenario. Even though the number of samples analysed is limited (<reflink idref="bib57" id="ref70">57</reflink>, 32 and 14 images for data set 1, 2, and 3 respectively), due to availability constraints, they can faithfully recapitulate the multifactoriality of the disease. In the following, we describe the statistical measures used to evaluate the automatic processing algorithms (Sectïon "Statistical measures"), show the DAB-brown signal quantification results (Section "Automatic recognition of DAB-brown signal"), present the results of detection (Section "Detection of nuclei") and classification (Section "Classification of nuclei") of nuclei and discuss the software experts' perception and performance (Section "Elapsed time and system usability"). Finally, we address the relevance of the obtained Cx43 measurements in the context of breast cancer biomedical research (Sectïon "Biomedical results").</p> <hd id="AN0175756719-12">Statistical measures</hd> <p>The statistical evaluation of the algorithms to automatically recognise the objects of interest (DAB-brown regions or nuclei) can be measured counting the number of hits and mistakes in the detection. We consider that a detection is a true positive (TP) hit when the user did not modify the automatic recognition provided by the algorithm; a false positive (FP) when the user manually deleted the objects provided by the computer; and a false negative (FN) when the user manually added a DAB region/nucleus.The number of DAB regions/nuclei automatically recognised is defined by the sensitivity (Se) or recall (R), positive predictivity value (PPV) or precision (P), average precision and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score, all in %:</p> <p>2 <ephtml> <math display="block" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mtable><mtr><mtd columnalign="right"><mstyle displaystyle="true" scriptlevel="0"><mrow><mi>S</mi><mi>e</mi><mo>=</mo><mi>R</mi><mo>=</mo><mn>100</mn><mfrac><mrow><mi mathvariant="italic">TP</mi></mrow><mrow><mi>F</mi><mi>N</mi><mo>+</mo><mi>T</mi><mi>P</mi></mrow></mfrac><mo>,</mo><mspace width="1em" /><mi>P</mi><mi>P</mi><mi>V</mi><mo>=</mo><mi>P</mi><mo>=</mo><mn>100</mn><mfrac><mrow><mi mathvariant="italic">TP</mi></mrow><mrow><mi>F</mi><mi>P</mi><mo>+</mo><mi>T</mi><mi>P</mi></mrow></mfrac><mo>,</mo><mspace width="1em" /><mi>A</mi><mi>P</mi><mo>=</mo><mn>100</mn><mfrac><mrow><mi mathvariant="italic">TP</mi></mrow><mrow><mi>T</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>P</mi><mo>+</mo><mi>F</mi><mi>N</mi></mrow></mfrac><mo>,</mo><mspace width="1em" /><msub><mi>F</mi><mn>1</mn></msub><mo>=</mo><mn>100</mn><mfrac><mrow><mn>2</mn><mi>P</mi><mi>R</mi></mrow><mrow><mi>P</mi><mo>+</mo><mi>R</mi></mrow></mfrac></mrow></mstyle></mtd></mtr></mtable></mrow></math> </ephtml></p> <p>Graph</p> <p>In the case of positivity estimation, the BrownDetector algorithm can be evaluated measuring the area percentage occupied by DAB-brown regions reported by the computer (APC) and after the expert's supervision (APES). The difference <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mo stretchy="false">|</mo><mi>A</mi><mi>P</mi><mi>C</mi><mo>-</mo><mi>A</mi><mi>P</mi><mi>E</mi><mi>S</mi><mo stretchy="false">|</mo></mrow></math> </ephtml> provides the computer error for estimating the image positivity.</p> <p>The performance of the SVM model in the prediction of the nuclei category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi mathvariant="script">C</mi><mo>∈</mo></mrow></math> </ephtml> {highly stained, medium stained, low stained, no stain} is evaluated using the Cohen kappa, defined in Eq. 1 above, and the accuracy (in %), whose value is 100 multiplied by the number of nuclei correctly classified by the classifier ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>p</mi><mi>a</mi></msub></math> </ephtml> in the Eq. 1) and divided by the total number of nuclei. The <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>S</mi><msub><mi>e</mi><mi>i</mi></msub></mrow></math> </ephtml> and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>P</mi><mi>P</mi><msub><mi>V</mi><mi>i</mi></msub></mrow></math> </ephtml> of each category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>i</mi><mo>∈</mo><mi mathvariant="script">C</mi></mrow></math> </ephtml> are also calculated considering that: (<reflink idref="bib1" id="ref71">1</reflink>) the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><msub><mi>P</mi><mi>i</mi></msub></mrow></math> </ephtml> are the number of nuclei of category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>C</mi><mi>i</mi></msub></math> </ephtml> correctly classified by the SVM into the category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">C</mi><mi>i</mi></msub></math> </ephtml> ; (<reflink idref="bib2" id="ref72">2</reflink>) the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>F</mi><msub><mi>P</mi><mi>i</mi></msub></mrow></math> </ephtml> are the number of nuclei classified into category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">C</mi><mi>i</mi></msub></math> </ephtml> , but whose true category label is other; (<reflink idref="bib3" id="ref73">3</reflink>) the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>F</mi><msub><mi>N</mi><mi>i</mi></msub></mrow></math> </ephtml> are the number of nuclei of true category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">C</mi><mi>i</mi></msub></math> </ephtml> that the classifier assigned to other category; and (<reflink idref="bib4" id="ref74">4</reflink>) <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>T</mi><msub><mi>N</mi><mi>i</mi></msub></mrow></math> </ephtml> (True Negative) is the number of nuclei of true category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">C</mi><mi>j</mi></msub></math> </ephtml> , <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>j</mi><mo>≠</mo><mi>i</mi></mrow></math> </ephtml> , classified by SVM as any category <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi mathvariant="script">C</mi><mi>j</mi></msub></math> </ephtml> , <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>j</mi><mo>≠</mo><mi>i</mi></mrow></math> </ephtml> . The specificity ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mi>S</mi><msub><mi>p</mi><mi>i</mi></msub></mrow></math> </ephtml> ) is defined as:</p> <p>3 <ephtml> <math display="block" xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mtable><mtr><mtd columnalign="right"><mstyle displaystyle="true" scriptlevel="0"><mrow><mi>S</mi><msub><mi>p</mi><mi>i</mi></msub><mo>=</mo><mn>100</mn><mfrac><mrow><mi>T</mi><msub><mi>N</mi><mi>i</mi></msub></mrow><mrow><mi>T</mi><msub><mi>N</mi><mi>i</mi></msub><mo>+</mo><mi>F</mi><msub><mi>P</mi><mi>i</mi></msub></mrow></mfrac></mrow></mstyle></mtd></mtr></mtable></mrow></math> </ephtml></p> <p>Graph</p> <hd id="AN0175756719-13">Automatic recognition of DAB-brown signal</hd> <p>Table 1 shows the results of the BrownDetector algorithm (BrownDetector algorithm to recognise DAB-brown signal section) to automatically recognise the DAB-brown regions in the IHC images, while APC and APES are the area percentages automatically detected by the computer and after the expert's supervision, respectively. In all data sets, the sensitivity for recognizing DAB-brown areas is higher than 99% and the positive predictivity value is higher than 95%. Hence, the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score achieves values higher than 96% for all data sets. The difference between the area percentage recognised automatically by the computer and after expert's supervision is only 0.07 and 0.05 points for SET 1 and SET 2 respectively. Hence, the experts only supervise the 10% and 28% of images for SET 1 and SET 2 respectively. But, it is important to emphasise that the maximum difference of area percentage was only of 1.5 and 1.67 points for SET 1 and 2 respectively, i.e. the area of the DAB-brown regions corrected by experts was very low and affects to very small regions. For the SET 3, the software operates perfectly (precision of 100%). This good performance of the automatic BrownDetector algorithm leads us to conclude that the system could be used in routinary tasks of biomedical labs.</p> <p>Table 1 Sensitivity (Se), positive predictivity value (PPV), average precision (AP) and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score, in %, of BreastAnalyser working in the lab to recognise the DAB-brown regions in IHC images. The APC and APES are the area percentages detected automatically by the computer and after the expert's supervision, respectively.</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left"><p>Data set</p></th><th align="left"><p>#images</p></th><th align="left"><p>Se</p></th><th align="left"><p>PPV</p></th><th align="left"><p>AP</p></th><th align="left"><p><math xmlns="http://www.w3.org/1998/Math/MathML"><msub xmlns=""><mi>F</mi><mn>1</mn></msub></math><inline-graphic href="41598_2024_53002_Article_IEq162.gif" /></p></th><th align="left"><p>APC</p></th><th align="left"><p>APES</p></th><th align="left"><p>|APC-APES|</p></th></tr></thead><tbody><tr><td align="left"><p><bold>SET 1</bold></p></td><td align="left"><p>47</p></td><td align="left"><p>99.2</p></td><td align="left"><p>95.5</p></td><td align="left"><p>95.3</p></td><td align="left"><p>96.1</p></td><td align="left"><p>9.13</p></td><td align="left"><p>9.06</p></td><td align="left"><p>0.07</p></td></tr><tr><td align="left"><p><bold>SET 2</bold></p></td><td align="left"><p>32</p></td><td align="left"><p>99.8</p></td><td align="left"><p>95.4</p></td><td align="left"><p>95.1</p></td><td align="left"><p>97.2</p></td><td align="left"><p>2.68</p></td><td align="left"><p>2.66</p></td><td align="left"><p>0.05</p></td></tr><tr><td align="left"><p><bold>SET 3</bold></p></td><td align="left"><p>14</p></td><td align="left"><p>100.0</p></td><td align="left"><p>100.0</p></td><td align="left"><p>100.0</p></td><td align="left"><p>100</p></td><td align="left"><p>22.92</p></td><td align="left"><p>22.92</p></td><td align="left"><p>0.00</p></td></tr></tbody></table> </ephtml> </p> <hd id="AN0175756719-14">Detection of nuclei</hd> <p>Table 2 shows the statistical results provided by RBA and EBA (see "RBA to detect the nuclei" and "EBA algorithm to detect the nuclei" sections, respectively) to automatically detect the nuclei in IHC images. The sensitivity is quite different among sets (70.4% and 48.2% for SET 3 and 1 respectively and only 16.6% for SET 2), which means that the algorithms detect much more true nuclei in the SET 3. Nevertheless, the positive predictivity value is quite similar and high for all sets (about 70%), denoting that the algorithms do no spot many false positive nuclei. The average precision includes both types of errors, misdetections and false detections, but it is dominated by the sensitivity behaviour (63.8% for SET 3, 41.4% for SET 1 and 14.7% for SET 2). This behaviour is also observed in <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score, which ranged between 24.1% for SET 2 and 96.1% for SET 3. Figure 8 shows the <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score for nuclei detection for all images. Although the nuclei detection in SET 3 is much higher than in the remaining data sets, a high variability among images can be observed in all data sets, ranging from 0% (i.e. the algorithm does not detect nuclei or detects many false nuclei, so a better option is to manually detect the nuclei using a mouse click) up to <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> =96%. The nuclei detection algorithms work rather poorly when the nuclei are partially masked by the DAB-brown signal.</p> <p>Table 2 Sensitivity (Se), positive predictivity value (PPV), average precision (AP) and <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score, in %, of BreastAnalyser for the nuclei detection on each set. The value N reports the average number of nuclei per ROI or image.</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left"><p>Data set</p></th><th align="left"><p>#images</p></th><th align="left"><p>#ROI</p></th><th align="left"><p>N</p></th><th align="left"><p>Se</p></th><th align="left"><p>PPV</p></th><th align="left"><p>AP</p></th><th align="left"><p><math xmlns="http://www.w3.org/1998/Math/MathML"><msub xmlns=""><mi>F</mi><mn>1</mn></msub></math><inline-graphic href="41598_2024_53002_Article_IEq167.gif" /></p></th></tr></thead><tbody><tr><td align="left"><p><bold>SET 1</bold></p></td><td align="left"><p>47</p></td><td align="left"><p>58</p></td><td align="left"><p>55</p></td><td align="left"><p>48.2</p></td><td align="left"><p>71.8</p></td><td align="left"><p>41.4</p></td><td align="left"><p>55.2</p></td></tr><tr><td align="left"><p><bold>SET 2</bold></p></td><td align="left"><p>32</p></td><td align="left"><p>32</p></td><td align="left"><p>50</p></td><td align="left"><p>16.4</p></td><td align="left"><p>66.2</p></td><td align="left"><p>14.7</p></td><td align="left"><p>24.1</p></td></tr><tr><td align="left"><p><bold>SET 3</bold></p></td><td align="left"><p>14</p></td><td align="left"><p>13</p></td><td align="left"><p>45</p></td><td align="left"><p>70.4</p></td><td align="left"><p>76.5</p></td><td align="left"><p>63.8</p></td><td align="left"><p>96.1</p></td></tr></tbody></table> </ephtml> </p> <p>Graph: Figure 8 F1 -score (in %) for each image in SET 1 (green diamonds), SET 2 (red circles) and SET 3 (blue squares).</p> <hd id="AN0175756719-15">Classification of nuclei</hd> <p>In order to evaluate the classification of the nuclei according to their level of staining (<bold>high</bold>, <bold>medium</bold>, <bold>low</bold> and <bold>no</bold> staining), an expert assigned a category to all the nuclei in the ROIs studied and, subsequently, the expert's labels were compared with the labels provided by the classifier. Then, the default pre-trained classifier included in the BreastAnalyser software was used to predict the category label. The classification accuracy and kappa evaluate the agreement between both category labelings. In <bold>SET 1</bold>, accuracy and kappa are 48.9% and 18.2%, respectively. The confusion matrix is reported by Table 3. The value in row <emph>i</emph> and column <emph>j</emph> is <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>100</mn><msub><mi>N</mi><mrow><mi mathvariant="italic">ij</mi></mrow></msub><mo stretchy="false">/</mo><mi>N</mi></mrow></math> </ephtml> , where <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>N</mi><mrow><mi mathvariant="italic">ij</mi></mrow></msub></math> </ephtml> and <emph>N</emph> are the same as in Eq. 1 above. The diagonal numbers (in bold) give the percentage of nuclei correctly classified for each category, and the sum of the diagonal gives the classification accuracy. In general, the percentages in the main diagonal are higher than the percentages outside it, meaning that the classifier can discriminate the categories. Nevertheless, for some categories (e.g. low and no stained) the percentages are quite similar, which explains the relatively low values of accuracy, sensitivity and predictive positive value. The specificity is quite high for high and medium stained (above 92%) and moderated for low and no stain (about 60%). The staining level of the nuclei is a continuous property, so it is possible that the subjectivity of the experts or the non-linear responses of the human visual system label the nuclei depending on their surroundings. But, it is important to note that the classification is almost always confused with the neighbouring category (i.e., low and no staining). Hence, if we consider as success when the classifier predicts the true or the contiguous category, the classification accuracy is increased up to 95.59%. In line with this, if our main objective in the biomedical part of this work is to discriminate between weak and strong staining, as is customary in the literature, and if we consider only two classes (weak=no+low, strong=medium+high), then the accuracy is 89.32%.</p> <p>Table 3 Confusion matrix (in %) for nuclei staining classification (high, medium, low and without staining) for <bold>SET 1</bold>. The columns Se, PPV and Sp report the sensitivity, predictive positive value and specificity for each class, respectively.</p> <p> <ephtml> <table frame="hsides" rules="groups"><thead><tr><th align="left" /><th align="left" /><th align="left" /><th align="left" colspan="4"><p>Predicted category</p></th><th align="left" /><th align="left" /><th align="left" /></tr><tr><th align="left" /><th align="left" /><th align="left"><p>#Nucleus</p></th><th align="left"><p>High</p></th><th align="left"><p>Medium</p></th><th align="left"><p>Low</p></th><th align="left"><p>No</p></th><th align="left"><p>Se</p></th><th align="left"><p>PPV</p></th><th align="left"><p>Sp</p></th></tr></thead><tbody><tr><td align="left" /><td align="left"><p>High</p></td><td align="left"><p>51</p></td><td align="left"><p><bold>0.94</bold></p></td><td align="left"><p>0.54</p></td><td align="left"><p>0.07</p></td><td align="left"><p>0.17</p></td><td align="left"><p>54.90</p></td><td align="left"><p>34.15</p></td><td align="left"><p>98.2</p></td></tr><tr><td align="left"><p>True</p></td><td align="left"><p>Medium</p></td><td align="left"><p>327</p></td><td align="left"><p>1.27</p></td><td align="left"><p><bold>5.53</bold></p></td><td align="left"><p>1.81</p></td><td align="left"><p>2.35</p></td><td align="left"><p>50.46</p></td><td align="left"><p>46.87</p></td><td align="left"><p>92.9</p></td></tr><tr><td align="left"><p>Category</p></td><td align="left"><p>Low</p></td><td align="left"><p>1638</p></td><td align="left"><p>0.50</p></td><td align="left"><p>4.33</p></td><td align="left"><p><bold>23.47</bold></p></td><td align="left"><p>26.63</p></td><td align="left"><p>42.73</p></td><td align="left"><p>62.72</p></td><td align="left"><p>69.0</p></td></tr><tr><td align="left" /><td align="left"><p>No</p></td><td align="left"><p>967</p></td><td align="left"><p>0.03</p></td><td align="left"><p>1.41</p></td><td align="left"><p>12.07</p></td><td align="left"><p><bold>18.88</bold></p></td><td align="left"><p>58.28</p></td><td align="left"><p>39.31</p></td><td align="left"><p>57.0</p></td></tr></tbody></table> </ephtml> </p> <p>In the <bold>SETS 2</bold> and <bold>3</bold> the classification was evaluated within the routinary use of BreastAnalyser. After drawing the ROIs, the automatic detection of nuclei is executed and classification is performed on the nuclei detected automatically. Afterwards, the expert supervised jointly the detection and classification results. In these experiments, the accuracy was 100% for the SET 2 and 93% for SET 3. This accurate classification of the staining level can be explained by two reasons: (<reflink idref="bib1" id="ref75">1</reflink>) the expert did not change the nuclei labels; and (<reflink idref="bib2" id="ref76">2</reflink>) the correctly detected nuclei are not partially masked by other types of tissues, and then, in these cases, the classifier predicts correctly the true category.</p> <hd id="AN0175756719-16">Elapsed time and system usability</hd> <p>The elapsed time for the analysis of an IHC image depends mainly on the time required to review the automatic recognition of DAB-brown areas and the nuclei detection. This values were estimated by the INIBIC lab experts using a stopwatch to measure the time needed to load, process, review and save the results for each image on a standard personal computer with 4 cores AMD Ryzen 5 3500U at 2.10 GHz and 8 GB of RAM memory under Windows operative system. The average elapsed time estimated was 3.75 minutes. Obviously, the analysis time is dominated by the review time. So, in order to estimate the time spent by the automatic processing, additional experiments were done on a computer with 8 Intel® Core <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msup><mrow /><mrow><mi mathvariant="italic">TM</mi></mrow></msup></math> </ephtml> i7-9700K CPUs at 3.60 GHz and 64 GB of RAM memory under Linux Kubuntu 20.04. The average time spent to process an image of <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>2040</mn><mo>×</mo><mn>1436</mn></mrow></math> </ephtml> pixels was 0.43 miliseconds to BrownDetector algorithm, 0.49 miliseconds to EBA and 0.16 miliseconds to RBA.</p> <p>The subjective expert perception and system usability was evaluated using the System Usability Scale (SUS) questionnaire[<reflink idref="bib43" id="ref77">43</reflink>], which measures the learning abitity and perceived usability of the software. The SUS has 10 items with a five-point scale. If the score is below 25 it is the worst imaginable system; between 25 and 39 is "from worst imaginable to poor"; between 40 and 52 is "from poor to OK"; from 52 to 73 is "OK to good"; from 73 to 85 is "good to excellent" and from 85 to 100 is "excellent to best imaginable"[<reflink idref="bib44" id="ref78">44</reflink>]. Ten experts from the participating biomedical laboratories belonging to clinical pathology and biomedical research areas filled out the questionnaire and gave an average score of 85.5 points, ranging from 67 to 97.5 points. This result means that BreastAnalyser is perceived as excellent.</p> <hd id="AN0175756719-17">Biomedical results</hd> <p>BreastAnalyser was used to quantitate Cx43 immunostaining in various invasive breast cancer tissue samples, representative of the complexity and inter-patient variability of clinical specimens. This study was performed for the first two data sets: (<reflink idref="bib1" id="ref79">1</reflink>) <bold>SET 1</bold> to analyse potential variations in Cx43 expression depending on the breast cancer subtype; and (<reflink idref="bib2" id="ref80">2</reflink>) <bold>SET 2</bold> to discern possible differences in Cx43 levels in invasive breast cancer luminal B HER2- tissues according to their Oncotype DX Breast Recurrence Score Test results (Figs. 1 and 2 show representative IHC images). All images of <bold>SET 1</bold> and <bold>SET 2</bold> were automatically analysed and supervised, semi-automatically, by an histopathological expert using BreastAnalyser.</p> <p>For <bold>SET 1</bold>, Fig. 9 shows bar charts representing Cx43 expression in core needle biopsies of invasive tumours not subjected to chemotherapy or radiotherapy and healthy control tissues, quantitated as brown area in pixels relativized to cell density for each patient. According to these measurements, all breast tumour samples showed significantly lower Cx43 immunostaining than healthy mammary controls. Furthermore, Cx43 signal seemed to be inversely and significantly correlated with breast cancer subtype aggressiveness, being the lowest in the most serious subtype triple negative (ER, PR and HER2 negative), slightly higher in non-luminal HER2+ (ER and PR negative, HER2 positive) and luminal B HER2- (ER and/or PR positive, HER2 negative), and the highest in luminal B HER2+ (ER and/or PR positive, HER2 positive) and in the better-prognosis luminal A (ER and/or PR positive, HER2 negative). In line with these results, Cx43 and/or GJ function downregulation has been extensively described in breast cancer cell lines[<reflink idref="bib27" id="ref81">27</reflink>],[<reflink idref="bib28" id="ref82">28</reflink>] and primary tumour tissues, suggesting a tumour-suppressing role for Cx43 in primary breast tumours[<reflink idref="bib27" id="ref83">27</reflink>],[<reflink idref="bib31" id="ref84">31</reflink>]. Several studies have also appointed a positive correlation between Cx43 gene and protein levels and ER and PR status, and a negative relationship with HER2 expression[<reflink idref="bib30" id="ref85">30</reflink>],[<reflink idref="bib45" id="ref86">45</reflink>]. In our work, no significant difference was observed regarding Cx43 immunostaining between HER2+ and triple negative breast cancer, as also reported in the literature[<reflink idref="bib29" id="ref87">29</reflink>],[<reflink idref="bib46" id="ref88">46</reflink>].</p> <p>Graph: Figure 9Cx43 is progressively downregulated with increasing breast cancer subtype malignancy. Graph represents Cx43 expression (quantitated as brown area in pixels) relativized to cell density for each patient. Mean+SD; mean values specified in the graph. n=5-8/subtype, each n is represented as a dot in the graph. One-way ANOVA. * P<0.05 , ** P<0.01 , *** P<0.001.</p> <p>Graph: Figure 10Higher Cx43 is associated to lower cancer recurrence risk in Oncotype DX-tested luminal B HER2- samples. Left graph represents Cx43 expression (quantitated as brown area in pixels) relativized to cell density for each patient. Right graph shows the percentage of cellular nuclei with high, medium, low or no Cx43 nuclear signal. Mean+SD; mean values specified in the graph. n=6/category, each n is represented as a dot in the graph. T-test. ** P<0.01 , *** P<0.001.</p> <p>Next, this analysis was mirrored with images from <bold>SET 2</bold>, comprising invasive breast cancer luminal B HER2- tissues assessed for Oncotype DX Breast Recurrence Score Test. Indeed, Cx43 expression was significantly higher in low cancer recurrence risk tissues than in high risk ones (Fig. 10 left). In order to move one step forward in analysing Cx43 levels in these samples, nuclear Cx43 signal was compared among them, categorizing the cellular nuclei in high, medium, low or without Cx43 staining (Fig. 10 right). The majority of the samples were characterized by null ("without" category) or almost null ("low" category) nuclear Cx43 signal, with high-risk patients having significantly more Cx43-null nuclei than low-risk ones (95% versus 84.9%, respectively). As for nuclei with low Cx43 levels, they were significantly more abundant in low-risk tissues than in high-risk samples (13.2% versus 4.1%, respectively). Regarding nuclei with medium Cx43 staining, low-risk patients tended to present more than high-risk ones, although the difference was not significant (1.8% versus 0.1%, respectively, Fig. 10 right). It became apparent that nuclear Cx43 levels mimicked those of the whole tissue (Fig. 10), with low-risk samples having significantly higher Cx43/more Cx43-positive nuclei (even with low staining levels) than high-risk tissues. In agreement with these data, higher Cx43 transcript and protein levels have previously been significantly associated with better prognosis in breast cancer, namely high overall and relapse-free survival and lower disease recurrence, further reinforcing the potential prognostic value of Cx43 in breast carcinoma[<reflink idref="bib29" id="ref89">29</reflink>],[<reflink idref="bib30" id="ref90">30</reflink>],[<reflink idref="bib47" id="ref91">47</reflink>],[<reflink idref="bib48" id="ref92">48</reflink>]. It is key to note that our study is the first to validate BreastAnalyser and to address Cx43 expression in the context of breast cancer samples scored according to the pioneering diagnostic test Oncotype DX, putting forward the possibility of Cx43 as a complementary prognostic marker to take into consideration.</p> <p>Finally, regarding images from <bold>SET 3</bold>, explicitly included just for the assessment of inter-laboratory software performance, brown signal was subsequently not evaluated at the biomedical level. Nevertheless, it is relevant to note that they present different magnifications and belong to antigens routinely assessed in breast cancer clinical environment. The positive outcome and efficiency of Breast Analyser in this context presents this software as a versatile tool which can also potentially support anatomo- and histo-pathological clinical decisions.</p> <hd id="AN0175756719-18">Conclusions</hd> <p>Breast cancer is the most diagnosed cancer worldwide and represents the fifth cause of cancer mortality globally. Immunohistochemistry can support the oncological diagnosis, therapeutic decisions and biomarker discovery but, currently, its evaluation is often subjective and qualitative due to the lack of suitable image analysis tools. It is also known the heterogeneity of IHC images, intimately related to the high complexity of breast cancer and to the inherent intricacy of pathological images due to differences in tissue processing, staining, image acquisition, etc. The available software tools can be grouped into: (<reflink idref="bib1" id="ref93">1</reflink>) generic free tools to analyse the images, in which the user needs some programming skills for an optimal use; and (<reflink idref="bib2" id="ref94">2</reflink>) sophisticated algorithms that perform specific task, but are often very time consuming. Some drawbacks of both approaches are the lack of suitable tools to review the analysis before the IHC image quantification, or the need of programming and image analysis knowledge for an optimal use.</p> <p>Our software BreastAnalyser combines the automatic processing of the IHC image using sophisticated algorithms with a friendly GUI that allows experts to review the analysis before image quantification. The segmentation algorithms implicitly include image variability, saving normalization time. The recognition of DAB-brown signal is almost perfect for all data sets tested (sensitivities higher than 99% and positive predictivities above 95%). The detection of nuclei achieves lower performances strongly depending on the image ( <ephtml> <math xmlns="http://www.w3.org/1998/Math/MathML"><msub><mi>F</mi><mn>1</mn></msub></math> </ephtml> -score from 0 to 96% for different images). Nuclei classification according to their staining level into categories high, medium, low and without staining is very good (accuracy higher than 93% for SET 2 and SET 1). The elapsed time to automatically process the IHC images was less than one second, and the analysis time of an expert per image depends on the review requirements, but it can be estimated in 3.7 minutes per image, spending only 0.43 ms for brown detection and 0.49 or 0.16 for nuclei detection depending on the method (EBA or RBA) used. The overall perception of BreastAnalyser using the SUS questionnaire reported an average of 85.5 points, which means that the system is perceived as excellent.</p> <p>Within the biomedical context, BreastAnalyser allowed to perform quantifications of the breast cancer IHC images, leading to the following conclusions: (<reflink idref="bib1" id="ref95">1</reflink>) Cx43 is deeply downregulated at the protein level in human invasive breast cancer tissue samples when compared to normal breast, with a tendency to decrease its expression as the subtype malignancy increases. It is minimal in triple negative breast cancer (the most aggressive), followed by HER2+, luminal B HER2-, luminal B HER2+ and finally luminal A (the least aggressive) breast cancer, with the highest levels; and (<reflink idref="bib2" id="ref96">2</reflink>) higher Cx43 protein expression is associated to lower cancer recurrence risk in Oncotype DX-tested luminal B HER2- breast cancer tissues. Nuclear Cx43 levels in these samples mimicked those of the whole tissue, with low-risk samples having significantly higher Cx43/more Cx43-positive nuclei (although marginally stained) than high-risk tissues.</p> <p>BreastAnalyser software can competently support both basic and translational research, offering a straightforward and reliable approach for IHC analysis that overcomes some other available softwares in terms of simplicity and pragmatism. It might also be advantageous for certain tasks of pathological anatomy research, namely automatization of IHC-brown signal quantitation, presenting a potential clinical tool for breast cancer diagnosis and prognosis. Future work includes the validation of BreastAnalyser in more biomedical labs and to test it to quantify positivity on IHC images of other tissues, for which our preliminary visual experiments are very encouraging. From the image segmentation point of view, future work will focus on the improvement of automatic nuclei detection and stratification.</p> <hd id="AN0175756719-19">Ethical approval statement</hd> <p>The present study was conducted following the guidelines of the Declaration of Helsinki and later amendments, and was approved by the regional research ethics committees of the Autonomous Community of Galicia-CAEIG (registry codes 2019/535 and 2015/029, INIBIC lab samples) and the Albacete Integrated Care Management System (registry code 2020/06/071, CRIB lab samples). All tissue donors were adults and informed consent was obtained from them.</p> <hd id="AN0175756719-20">Supplementary Material</hd> <p>The user guide of the BreastAnalyser software is included as a supplementary material.</p> <hd id="AN0175756719-21">Acknowledgements</hd> <p>This work has received financial support from Xunta de Galicia (Centro singular de investigación de Galicia, accreditation 2020-2023, to E.C. and M.F-D.), the European Union (European Regional Development Fund, to E.C. and M.F-D.), Health Institute Carlos III (ISCIII)-the European Regional Development Fund "A way of making Europe" from the European Union (PI19/00145, to M.D.M., PI22/00025 to M.F-D.), Xunta de Galicia-GAIN (IN852D and IN607B2020/12, to M.D.M.), the European Union (HORIZON-CSA 101079489 TWINFLAG, to M.D.M.) and the Spanish Ministry of Education (FPU15/04237 to M.R-C.M.). We acknowledge Angel Concha and the A Coruña Biobank for some of the breast tissue samples.</p> <hd id="AN0175756719-22">Author contributions</hd> <p>M.R.C.-M., M.A., E.C. and M.D.M. designed and performed the experiments, analysed the data and prepared the figures. P.S.-F., E.M.G.-M. M.F.-D., R.B.A., J.M. and B.A. contributed with clinical and technical advice and provided samples and validation assays. E.C. and M.D.M. conceived, directed and supervised the study. M.R.C.-M., M.A., E.C. and M.D.M wrote the manuscript with input from all co-authors. All authors reviewed and approved the manuscript.</p> <hd id="AN0175756719-23">Data availability</hd> <p>BreastAnalyser software and the data sets analysed and annotated during the current study are available in the CiTIUS repository https://citius.usc.es/transferencia/software/breastanalyser.</p> <hd id="AN0175756719-24">Competing interests</hd> <p>The authors declare no competing interests.</p> <hd id="AN0175756719-25">Supplementary Information</hd> <p>Graph: Supplementary Information.</p> <hd id="AN0175756719-26">Supplementary Information</hd> <p>The online version contains supplementary material available at https://doi.org/10.1038/s41598-024-53002-6.</p> <hd id="AN0175756719-27">Publisher's note</hd> <p>Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.</p> <ref id="AN0175756719-28"> <title> References </title> <blist> <bibl id="bib1" idref="ref1" type="bt">1</bibl> <bibtext> Ferlay, et al. An overview. Int. J. Cancer. https://doi.org/10.1016/j.media.2022.102486 (2020).</bibtext> </blist> <blist> <bibl id="bib2" idref="ref2" type="bt">2</bibl> <bibtext> Sung H. Global cancer statistics 2020: Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J. Clin. 2021; 71: 209-249. 10.3322/caac.21660. 33538338</bibtext> </blist> <blist> <bibl id="bib3" idref="ref3" type="bt">3</bibl> <bibtext> Perou CM. Molecular portraits of human breast tumours. Nature. 2000; 406: 747-752. 2000Natur.406.747P. 10.1038/35021093. 10963602</bibtext> </blist> <blist> <bibl id="bib4" idref="ref4" type="bt">4</bibl> <bibtext> Sørlie T. Gene expression patterns of breast carcinomas distinguish tumor subclasses with clinical implications. Proc. Natl. Acad. Sci. 2001; 98: 10869-74. 2001PNAS.9810869S. 10.1073/pnas.191367098. 11553815. 58566</bibtext> </blist> <blist> <bibl id="bib5" idref="ref6" type="bt">5</bibl> <bibtext> Dai X. Breast cancer intrinsic subtype classification, clinical use and future trends. Am. J. Cancer Res. 2015; 5; 10: 2929-43. 26693050. 4656721</bibtext> </blist> <blist> <bibl id="bib6" idref="ref7" type="bt">6</bibl> <bibtext> Slembrouck L. Concordance between results of inexpensive statistical models and multigene signatures in patients with ER+/HER2- early breast cancer. Modern Pathol. 2021; 34: 1297-1309. 10.1038/s41379-021-00743-8</bibtext> </blist> <blist> <bibl id="bib7" idref="ref8" type="bt">7</bibl> <bibtext> Ross JS, Hatzis C, Symmans WF, Pusztai L, Hortobágyi GN. Commercialized multigene predictors of clinical outcome for breast cancer. Oncologist. 2008; 13: 477-493. 10.1634/theoncologist.2007-0248. 18515733</bibtext> </blist> <blist> <bibl id="bib8" idref="ref41" type="bt">8</bibl> <bibtext> Yao K, Tong C-Y, Cheng C. A framework to predict the applicability of Oncotype DX, MammaPrint, and E2F4 gene signatures for improving breast cancer prognostic prediction. Sci. Rep. 2022; 12: 2211. 2022NatSR.12.2211Y. 10.1038/s41598-022-06230-7. 35140308. 8828770</bibtext> </blist> <blist> <bibl id="bib9" idref="ref9" type="bt">9</bibl> <bibtext> Syed YY. Oncotype DX breast recurrence score®: A review of its use in early-stage breast cancer. Mole. Diagn. Therapy. 2020; 24: 621-632. 10.1007/s40291-020-00482-7</bibtext> </blist> <blist> <bibtext> Jaffer S, Bleiweiss IJ. Beyond hematoxylin and eosin-the role of immunohistochemistry in surgical pathology. Cancer Invest. 2004; 22: 445-465. 10.1081/CNV-200034896. 15493365</bibtext> </blist> <blist> <bibtext> de Matos LL, Trufelli DC, de Matos MGL, da Silva-Pinhal MA. Immunohistochemistry as an important tool in biomarkers detection and clinical practice. Biomark Insights. 2010; 5: 9-20. 10.4137/bmi.s2185. 20212918. 2832341</bibtext> </blist> <blist> <bibtext> Walker RA. Quantification of immunohistochemistry-issues concerning methods, utility and semiquantitative assessment i. Histopathology. 2006; 49: 406-410. 10.1111/j.1365-2559.2006.02514.x. 16978204</bibtext> </blist> <blist> <bibtext> Rueden, C. T. et al. ImageJ2: ImageJ for the next generation of scientific image data. BMC Bioinf18 (2017).</bibtext> </blist> <blist> <bibtext> Bankhead, P. et al. Qupath: Open source software for digital pathology image analysis. Sci Reports7 (2017).</bibtext> </blist> <blist> <bibtext> Lu W. SlideGraph+: Whole slide image level graphs to predict HER2 status in breast cancer. Med. Imag. Anal. 2022; 80. 10.1016/j.media.2022.102486</bibtext> </blist> <blist> <bibtext> Qaiser T, Rajpoot NM. Learning where to see: A novel attention model for automated immunohistochemical scoring. IEEE Trans. Med. Imag. 2019; 38: 2620-2631. 10.1109/TMI.2019.2907049</bibtext> </blist> <blist> <bibtext> Liu J. An end-to-end deep learning histochemical scoring system for breast cancer TMA. IEEE Trans. Med. Imag. 2019; 38: 617-628. 10.1109/TMI.2018.2868333</bibtext> </blist> <blist> <bibtext> Roszkowiak L. System for quantitative evaluation of DAB & H-stained breast cancer biopsy digital images (CHISEL). Sci. Rep. 2021; 11: 9291. 2021NatSR.11.9291R. 10.1038/s41598-021-88611-y. 33927266. 8085130</bibtext> </blist> <blist> <bibtext> Rizzardi AE. Quantitative comparison of immunohistochemical staining measured by digital image analysis versus pathologist visual scoring. Diagn. Pathol. 2012; 7: 42. 10.1186/1746-1596-7-42. 22515559. 3379953</bibtext> </blist> <blist> <bibtext> Abdolhoseini M, Kluge MG, Walker FR, Johnson SJ. Segmentation of heavily clustered nuclei from histopathological images. Sci. Rep. 2019; 9: 4551. 2019NatSR.9.4551A. 10.1038/s41598-019-38813-2. 30872619. 6418222</bibtext> </blist> <blist> <bibtext> Wang P, Hu X, Li Y, Liu Q, Zhu X. Automatic cell nuclei segmentation and classification of breast cancer histopathology images. Signal Process. 2016; 122: 1-13. 10.1016/j.sigpro.2015.11.011</bibtext> </blist> <blist> <bibtext> Cordido A, Cernadas E, Fernández-Delgado M, García-González MA. Cystanalyser: A new software tool for the automatic detection and quantification of cysts in polycystic kidney and liver disease, and other cystic disorders. PLoS Comput. Biol. 2020; 16: 1-18. 10.1371/journal.pcbi.1008337</bibtext> </blist> <blist> <bibtext> Mbaidin A. STERapp: Semiautomatic software for stereological analysis application in the estimation of fish fecundity. Electronics. 2021. 10.3390/electronics10121432</bibtext> </blist> <blist> <bibtext> Harris A. Connexin channel permeability to cytoplasmic molecules. Prog. Biophys. Mol. Biol. 2007; 94: 120-43. 10.1016/j.pbiomolbio.2007.03.011. 17470375. 1995164</bibtext> </blist> <blist> <bibtext> Hervé J-C, Bourmeyster N, Sarrouilhe D, Duffy HS. Gap junctional complexes: From partners to functions. Prog. Biophys. Mol. Biol. 2007; 94: 29-65. 10.1016/j.pbiomolbio.2007.03.010. 17507078</bibtext> </blist> <blist> <bibtext> Deen, N. N. A, Abouhaidar, M. G. & Talhouk, R. S. Connexin43 as a Tumor Suppressor: Proposed Connexin43 mRNA-circularRNAs-microRNAs Axis Towards Prevention and Early Detection in Breast Cancer. Front. Med.6 (2019).</bibtext> </blist> <blist> <bibtext> Laird DW. Deficiency of Connexin43 Gap Junctions Is an Independent Marker for Breast Tumors1. Can. Res. 1999; 59: 4104-4110</bibtext> </blist> <blist> <bibtext> Li Z, Zhou Z, Welch DR, Donahue HJ. Expressing connexin 43 in breast cancer cells reduces their metastasis to lungs. Clin. Exp. Metastasis. 2008; 25: 893-901. 10.1007/s10585-008-9208-9. 18839320. 2754227</bibtext> </blist> <blist> <bibtext> Busby M, Hallett MT, Plante I. The complex subtype-dependent role of connexin 43 (GJA1) in breast cancer. Int. J. Mol. Sci. 2018; 19: 693. 10.3390/ijms19030693. 29495625. 5877554</bibtext> </blist> <blist> <bibtext> Chasampalioti M. Connexin 43 is an independent predictor of patient outcome in breast cancer patients. Breast Cancer Res. Treat. 2019; 174: 93-102. 10.1007/s10549-018-5063-9. 30474779</bibtext> </blist> <blist> <bibtext> Kańczuga-Koda L. Expression of connexin 43 in breast cancer in comparison with mammary dysplasia and the normal mammary gland. Folia Morphol. 2003; 62: 439-42</bibtext> </blist> <blist> <bibtext> Sonka, M, Hlavac, V. & Boyle, R. Image Processing, Analysis, and Machine Vision (Thomson-Engineering, 2007).</bibtext> </blist> <blist> <bibtext> Azevedo-Tosta TA, de Faria PR, Neves LA, do Nascimento MZ. Computational normalization of H &E-stained histological images: Progress, challenges and future potential. Artif. Intel. Med. 2019; 95: 118-132. 10.1016/j.artmed.2018.10.004</bibtext> </blist> <blist> <bibtext> Cho H, Nishimura K, Watanabe K, Bise R. Effective pseudo-labeling based on heatmap for unsupervised domain adaptation in cell detection. Med. Imag. Anal. 2022; 79. 10.1016/j.media.2022.102436</bibtext> </blist> <blist> <bibtext> Cernadas E, Fernández-Delgado M, González-Rufino E, Carrión P. Influence of normalization and color space to color texture classification. Pattern Recogn. 2017; 61: 120-138. 2017PatRe.61.120C. 10.1016/j.patcog.2016.07.002</bibtext> </blist> <blist> <bibtext> Otsu N. A threshold selection method from gray-level histograms. IEEE Trans. Syst. Man Cyber. 1979; 9: 62-66. 10.1109/TSMC.1979.4310076</bibtext> </blist> <blist> <bibtext> Suzuki S, Be K. Topological structural analysis of digitized binary images by border following. Comput. Vis. Graph. Image Proc. 1985; 30: 32-46. 10.1016/0734-189X(85)90016-7</bibtext> </blist> <blist> <bibtext> Canny JF. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell. 1986; 8: 679-698. 10.1109/TPAMI.1986.4767851. 21869365</bibtext> </blist> <blist> <bibtext> Pintor J. Govocitos: A software tool for estimating fish fecundity based on digital analysis of histological images. Comput. Electron. Agric. 2016; 125: 89-98. 10.1016/j.compag.2016.04.014</bibtext> </blist> <blist> <bibtext> Fernández-Delgado M, Cernadas E, Barro S, Amorim D. Do we need hundreds of classifiers to solve real world classification problems?. J. Mach. Learn. Res. 2014; 15: 3133-3181. 3277155</bibtext> </blist> <blist> <bibtext> Chang, C. & Lin, C. LIBSVM: a library for support vector machines. ACM T Intel. Syst. Tec.2, 27:1–27:27 (2011)A</bibtext> </blist> <blist> <bibtext> McHugh ML. Interrater reliability: The kappa statistic. Biochemia Medica. 2012; 22: 276-282. 10.11613/BM.2012.031. 23092060. 3900052</bibtext> </blist> <blist> <bibtext> Bangor A, Kortum P, Miller J. Determining what individual SUS scores mean: Adding and adjective rating scale. J. Usability Stud. 2009; 4: 114-123</bibtext> </blist> <blist> <bibtext> Brooke JSUS. A retrospective. J. Usability Stud. 2013; 8: 29-40</bibtext> </blist> <blist> <bibtext> Teleki I. The potential prognostic value of connexin 26 and 46 expression in neoadjuvant-treated breast cancer. BMC Cancer. 2013. 10.1186/1471-2407-13-50. 23374644. 3583680</bibtext> </blist> <blist> <bibtext> Fu Y. Hsa-miR-206 represses the proliferation and invasion of breast cancer cells by targeting Cx43. Eur. Rev. Med. Pharmacol. Sci. 2015; 19: 2091-2104. 26125274</bibtext> </blist> <blist> <bibtext> Teleki I. Correlations of differentially expressed gap junction connexins cx26, cx30, cx32, cx43 and cx46 with breast cancer progression and prognosis. PLoS ONE. 2014; 9: 1-12. 10.1371/journal.pone.0112541</bibtext> </blist> <blist> <bibtext> Wu D-P. Cx43 deficiency confers emt-mediated tamoxifen resistance to breast cancer via c-src/pi3k/akt pathway: Erratum. Int. J. Biol. Sci. 2022; 18: 5309-5311. 10.7150/ijbs.75891. 36147465. 9461667</bibtext> </blist> </ref> <aug> <p>By Marina Rodríguez-Candela Mateos; Maria Azmat; Paz Santiago-Freijanes; Eva María Galán-Moya; Manuel Fernández-Delgado; Rosa Barbella Aponte; Joaquín Mosquera; Benigno Acea; Eva Cernadas and María D. Mayán</p> <p>Reported by Author; Author; Author; Author; Author; Author; Author; Author; Author; Author</p> </aug> <nolink nlid="nl1" bibid="bib10" firstref="ref10"></nolink> <nolink nlid="nl2" bibid="bib11" firstref="ref11"></nolink> <nolink nlid="nl3" bibid="bib12" firstref="ref12"></nolink> <nolink nlid="nl4" bibid="bib13" firstref="ref13"></nolink> <nolink nlid="nl5" bibid="bib14" firstref="ref14"></nolink> <nolink nlid="nl6" bibid="bib15" firstref="ref15"></nolink> <nolink nlid="nl7" bibid="bib21" firstref="ref16"></nolink> <nolink nlid="nl8" bibid="bib22" firstref="ref17"></nolink> <nolink nlid="nl9" bibid="bib23" firstref="ref18"></nolink> <nolink nlid="nl10" bibid="bib24" firstref="ref19"></nolink> <nolink nlid="nl11" bibid="bib25" firstref="ref20"></nolink> <nolink nlid="nl12" bibid="bib26" firstref="ref21"></nolink> <nolink nlid="nl13" bibid="bib27" firstref="ref22"></nolink> <nolink nlid="nl14" bibid="bib28" firstref="ref23"></nolink> <nolink nlid="nl15" bibid="bib29" firstref="ref25"></nolink> <nolink nlid="nl16" bibid="bib31" firstref="ref26"></nolink> <nolink nlid="nl17" bibid="bib32" firstref="ref48"></nolink> <nolink nlid="nl18" bibid="bib33" firstref="ref49"></nolink> <nolink nlid="nl19" bibid="bib34" firstref="ref50"></nolink> <nolink nlid="nl20" bibid="bib35" firstref="ref51"></nolink> <nolink nlid="nl21" bibid="bib36" firstref="ref53"></nolink> <nolink nlid="nl22" bibid="bib37" firstref="ref55"></nolink> <nolink nlid="nl23" bibid="bib38" firstref="ref63"></nolink> <nolink nlid="nl24" bibid="bib39" firstref="ref64"></nolink> <nolink nlid="nl25" bibid="bib40" firstref="ref67"></nolink> <nolink nlid="nl26" bibid="bib41" firstref="ref68"></nolink> <nolink nlid="nl27" bibid="bib42" firstref="ref69"></nolink> <nolink nlid="nl28" bibid="bib57" firstref="ref70"></nolink> <nolink nlid="nl29" bibid="bib43" firstref="ref77"></nolink> <nolink nlid="nl30" bibid="bib44" firstref="ref78"></nolink> <nolink nlid="nl31" bibid="bib30" firstref="ref85"></nolink> <nolink nlid="nl32" bibid="bib45" firstref="ref86"></nolink> <nolink nlid="nl33" bibid="bib46" firstref="ref88"></nolink> <nolink nlid="nl34" bibid="bib47" firstref="ref91"></nolink> <nolink nlid="nl35" bibid="bib48" firstref="ref92"></nolink>
CustomLinks:
  – Url: https://resolver.ebsco.com/c/xy5jbn/result?sid=EBSCO:edsdoj&genre=article&issn=20452322&ISBN=&volume=14&issue=1&date=20240201&spage=1&pages=1-19&title=Scientific Reports&atitle=Software%20BreastAnalyser%20for%20the%20semi-automatic%20analysis%20of%20breast%20cancer%20immunohistochemical%20images&aulast=Marina%20Rodr%C3%ADguez-Candela%20Mateos&id=DOI:10.1038/s41598-024-53002-6
    Name: Full Text Finder (for New FTF UI) (s8985755)
    Category: fullText
    Text: Find It @ SCU Libraries
    MouseOverText: Find It @ SCU Libraries
  – Url: https://doaj.org/article/231a556edeec4e41961f4e706f43b3cc
    Name: EDS - DOAJ (s8985755)
    Category: fullText
    Text: View record from DOAJ
    MouseOverText: View record from DOAJ
Header DbId: edsdoj
DbLabel: Directory of Open Access Journals
An: edsdoj.231a556edeec4e41961f4e706f43b3cc
RelevancyScore: 1020
AccessLevel: 3
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 1020.39056396484
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Software BreastAnalyser for the semi-automatic analysis of breast cancer immunohistochemical images
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AR" term="%22Marina+Rodríguez-Candela+Mateos%22">Marina Rodríguez-Candela Mateos</searchLink><br /><searchLink fieldCode="AR" term="%22Maria+Azmat%22">Maria Azmat</searchLink><br /><searchLink fieldCode="AR" term="%22Paz+Santiago-Freijanes%22">Paz Santiago-Freijanes</searchLink><br /><searchLink fieldCode="AR" term="%22Eva+María+Galán-Moya%22">Eva María Galán-Moya</searchLink><br /><searchLink fieldCode="AR" term="%22Manuel+Fernández-Delgado%22">Manuel Fernández-Delgado</searchLink><br /><searchLink fieldCode="AR" term="%22Rosa+Barbella+Aponte%22">Rosa Barbella Aponte</searchLink><br /><searchLink fieldCode="AR" term="%22Joaquín+Mosquera%22">Joaquín Mosquera</searchLink><br /><searchLink fieldCode="AR" term="%22Benigno+Acea%22">Benigno Acea</searchLink><br /><searchLink fieldCode="AR" term="%22Eva+Cernadas%22">Eva Cernadas</searchLink><br /><searchLink fieldCode="AR" term="%22María+D%2E+Mayán%22">María D. Mayán</searchLink>
– Name: TitleSource
  Label: Source
  Group: Src
  Data: Scientific Reports, Vol 14, Iss 1, Pp 1-19 (2024)
– Name: Publisher
  Label: Publisher Information
  Group: PubInfo
  Data: Nature Portfolio, 2024.
– Name: DatePubCY
  Label: Publication Year
  Group: Date
  Data: 2024
– Name: Subset
  Label: Collection
  Group: HoldingsInfo
  Data: LCC:Medicine<br />LCC:Science
– Name: Subject
  Label: Subject Terms
  Group: Su
  Data: <searchLink fieldCode="DE" term="%22Medicine%22">Medicine</searchLink><br /><searchLink fieldCode="DE" term="%22Science%22">Science</searchLink>
– Name: Abstract
  Label: Description
  Group: Ab
  Data: Abstract Breast cancer is the most diagnosed cancer worldwide and represents the fifth cause of cancer mortality globally. It is a highly heterogeneous disease, that comprises various molecular subtypes, often diagnosed by immunohistochemistry. This technique is widely employed in basic, translational and pathological anatomy research, where it can support the oncological diagnosis, therapeutic decisions and biomarker discovery. Nevertheless, its evaluation is often qualitative, raising the need for accurate quantitation methodologies. We present the software BreastAnalyser, a valuable and reliable tool to automatically measure the area of 3,3’-diaminobenzidine tetrahydrocholoride (DAB)-brown-stained proteins detected by immunohistochemistry. BreastAnalyser also automatically counts cell nuclei and classifies them according to their DAB-brown-staining level. This is performed using sophisticated segmentation algorithms that consider intrinsic image variability and save image normalization time. BreastAnalyser has a clean, friendly and intuitive interface that allows to supervise the quantitations performed by the user, to annotate images and to unify the experts’ criteria. BreastAnalyser was validated in representative human breast cancer immunohistochemistry images detecting various antigens. According to the automatic processing, the DAB-brown area was almost perfectly recognized, being the average difference between true and computer DAB-brown percentage lower than 0.7 points for all sets. The detection of nuclei allowed proper cell density relativization of the brown signal for comparison purposes between the different patients. BreastAnalyser obtained a score of 85.5 using the system usability scale questionnaire, which means that the tool is perceived as excellent by the experts. In the biomedical context, the connexin43 (Cx43) protein was found to be significantly downregulated in human core needle invasive breast cancer samples when compared to normal breast, with a trend to decrease as the subtype malignancy increased. Higher Cx43 protein levels were significantly associated to lower cancer recurrence risk in Oncotype DX-tested luminal B HER2- breast cancer tissues. BreastAnalyser and the annotated images are publically available https://citius.usc.es/transferencia/software/breastanalyser for research purposes.
– Name: TypeDocument
  Label: Document Type
  Group: TypDoc
  Data: article
– Name: Format
  Label: File Description
  Group: SrcInfo
  Data: electronic resource
– Name: Language
  Label: Language
  Group: Lang
  Data: English
– Name: ISSN
  Label: ISSN
  Group: ISSN
  Data: 2045-2322
– Name: NoteTitleSource
  Label: Relation
  Group: SrcInfo
  Data: https://doaj.org/toc/2045-2322
– Name: DOI
  Label: DOI
  Group: ID
  Data: 10.1038/s41598-024-53002-6
– Name: URL
  Label: Access URL
  Group: URL
  Data: <link linkTarget="URL" linkTerm="https://doaj.org/article/231a556edeec4e41961f4e706f43b3cc" linkWindow="_blank">https://doaj.org/article/231a556edeec4e41961f4e706f43b3cc</link>
– Name: AN
  Label: Accession Number
  Group: ID
  Data: edsdoj.231a556edeec4e41961f4e706f43b3cc
PLink https://login.libproxy.scu.edu/login?url=https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&scope=site&db=edsdoj&AN=edsdoj.231a556edeec4e41961f4e706f43b3cc
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1038/s41598-024-53002-6
    Languages:
      – Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 19
        StartPage: 1
    Subjects:
      – SubjectFull: Medicine
        Type: general
      – SubjectFull: Science
        Type: general
    Titles:
      – TitleFull: Software BreastAnalyser for the semi-automatic analysis of breast cancer immunohistochemical images
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Marina Rodríguez-Candela Mateos
      – PersonEntity:
          Name:
            NameFull: Maria Azmat
      – PersonEntity:
          Name:
            NameFull: Paz Santiago-Freijanes
      – PersonEntity:
          Name:
            NameFull: Eva María Galán-Moya
      – PersonEntity:
          Name:
            NameFull: Manuel Fernández-Delgado
      – PersonEntity:
          Name:
            NameFull: Rosa Barbella Aponte
      – PersonEntity:
          Name:
            NameFull: Joaquín Mosquera
      – PersonEntity:
          Name:
            NameFull: Benigno Acea
      – PersonEntity:
          Name:
            NameFull: Eva Cernadas
      – PersonEntity:
          Name:
            NameFull: María D. Mayán
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 02
              Type: published
              Y: 2024
          Identifiers:
            – Type: issn-print
              Value: 20452322
          Numbering:
            – Type: volume
              Value: 14
            – Type: issue
              Value: 1
          Titles:
            – TitleFull: Scientific Reports
              Type: main
ResultId 1